Teaching translation in the age of generative AI: New paradigm, new learning?

Teaching translation in the age of generative AI New paradigm, new learning? Edited by JC Penet Joss Moorkens Masaru Yamada language science press Translation and Multilingual Natural Language Processing 25 Translation and Multilingual Natural Language Processing Editors: Oliver Czulo (Universität Leipzig), Silvia Hansen-Schirra (Johannes Gutenberg-Universität Mainz), Reinhard Rapp (Hochschule Magdeburg-Stendal), Mario Bisiada (Universitat Pompeu Fabra) In this series (see the complete series history at https://langsci-press.org/catalog/series/tmnlp): 14. Bisiada, Mario (ed.). Empirical studies in translation and discourse. 15. Tra&Co Group (ed.). Translation, interpreting, cognition: The way out of the box. 16. Nitzke, Jean & Silvia Hansen-Schirra. A short guide to post-editing. 17. Hoberg, Felix. Informationsintegration in mehrsprachigen Textchats: Der Skype Translator im Sprachenpaar Katalanisch-Deutsch. 18. Kenny, Dorothy (ed.). Machine translation for everyone: Empowering users in the age of artificial intelligence. (* 19. Kajzer-Wietrzny, Marta, Adriano Ferraresi, Ilmari Ivaska & Silvia Bernardini. Mediated discourse at the European Parliament: Empirical investigations. *) 20. Marzouk, Shaimaa. Sprachkontrolle im Spiegel der Maschinellen Übersetzung: Untersuchung zur Wechselwirkung ausgewählter Regeln der Kontrollierten Sprache mit verschiedenen Ansätzen der Maschinellen Übersetzung. 21. Frittella, Francesca Maria. Usability research for interpreter-centred technology: The case study of SmarTerp. 22. Prandi, Bianca. Computer-assisted simultaneous interpreting: A cognitive-experimental study on terminology. 23. Kenny, Dorothy (ed.). 机器翻译知识普及: 为人工智能时代的用户赋能. 24. Czulo, Oliver, Martin Kappus & Felix Hoberg (Hrsg.). Digitale Translatologie. 25. Penet, JC, Joss Moorkens & Masaru Yamada (eds.). Teaching translation in the age of generative AI: New paradigm, new learning? ISSN: 2364-8899 Teaching translation in the age of generative AI New paradigm, new learning? Edited by JC Penet Joss Moorkens Masaru Yamada language science press JC Penet, Joss Moorkens & Masaru Yamada (eds.). 2026. Teaching translation in the age of generative AI: New paradigm, new learning? (Translation and Multilingual Natural Language Processing 25). Berlin: Language Science Press. This title can be downloaded at: http://langsci-press.org/catalog/book/520 © 2026, the authors Published under the Creative Commons Attribution 4.0 Licence (CC BY 4.0): http://creativecommons.org/licenses/by/4.0/ ISBN: 978-3-96110-549-6 (Digital) 978-3-98554-169-0 (Hardcover) ISSN: 2364-8899 DOI: 10.5281/zenodo.17580856 Source code available from www.github.com/langsci/520 Errata: paperhive.org/documents/remote?type=langsci&id=520 Cover and concept of design: Ulrike Harbort Typesetting: Sebastian Nordhoff Proofreading: Matthew Korte Fonts: Libertinus, Arimo, DejaVu Sans Mono, Source Han Serif JA Typesetting software: XƎL A T EX Language Science Press Scharnweberstraße 10 10247 Berlin, Germany http://langsci-press.org [email protected] Storage and cataloguing done by FU Berlin Contents Introduction JC Penet iii Acknowledgements xi Abbreviations xiii I New paradigm: New skills & competences 1 Translation competence in the age of generative AI: Debates, dilemmas, directions Gary Massey & Maureen Ehrensberger-Dow 3 2 Generative AI as a facilitator of deliberate practice in translator training Erik Angelone 27 3 AI Literacy: The concept of suitability and core translation skills Ramon Inglada 49 II New paradigm: New knowledge 4 Teaching translation students about data in the age of generative AI Lynne Bowker 67 5 Teaching translation with AI: Bridging theory and practice through prompt engineering Masaru Yamada 87 6 Teaching AI ethics for translation students Joss Moorkens & Gökhan Doğru 105 JC Penet, Joss Moorkens & Masaru Yamada III New paradigm: New teaching approaches 7 Re-positioning the human translator as the future expert of GenAI translation in the translation classroom: The results of a collaborative study Senem Öner Bulut 125 8 Computer-assisted language mediation in teaching human-centred augmented translation Maria Zimina-Poirot 147 9 Teaching subtitling in the times of generative AI David Orrego-Carmona 167 10 AI in an L2 translation class Sonia Vandepitte 191 11 Gamification as a pedagogical instrument in interpreter training Sahar Othmani & Nermin Sharman 211 12 Embracing machine translation in L2 education: Bridging theory and practice in the AI Age Atsushi Mizumoto 233 Index 249 ii Introduction JC Penet Newcastle University, United Kingdom 1 Why this book? Since OpenAI launched ChatGPT back in November 2022, generative artificial intelligence (GenAI) has acted as a major disruptor in the translation industry and beyond. Of course, in any given industry disruptors are not intrinsically good or bad. They merely act as game changers, good and bad. Also commonly associated with “chatbots”, GenAI tools certainly are a game changer for the translation industry and, consequently, for translator education too. Using deep learning to train on vast corpora, not only do LLMs have the ability to “generate text which is often indistinguishable from text written by a human” (Moorkens et al. 2025: 188), but they can do much more than just translate texts. In this book, we shall therefore make a distinction between artificial intelligence (AI) and GenAI. In our context, AI refers broadly to all AI technologies, including neural machine translation (NMT). GenAI, however, refers to machine learning tools that generate media, including chat-based Large Language Models (LLMs) like GPT-4. Since this book deals mostly with text, the terms GenAI and LLMs tend to appear synonymously. Unlike freely available machine translation (MT) tools like Google Translate or DeepL, LLMs can also be prompted to adapt their translations to take the source text and/or the target text context into consideration – for free (initially). In a matter of seconds. It is easy to see how GenAI can potentially contribute to the further democratisation of translation by making it freely accessible in contexts where translation would have been neither practical nor affordable until now. This disruptor also provides translators with new resources that can potentially help them not just to increase the speed and quality of the translation work they deliver to clients, but also with the way they manage their translation projects or JC Penet. 2026. Introduction. In JC Penet, Joss Moorkens & Masaru Yamada (eds.), Teaching translation in the age of generative AI: New paradigm, new learning?, iii–ix. Berlin: Language Science Press. DOI: 10.5281/zenodo. 17641062 JC Penet go about finding new clients. In the 2024 edition of the annual ELIS (European Language Industry Survey) report, for instance, the independent translators who saw GenAI as a positive trend explained that they leveraged its possibilities “as a tool (e.g. terminology extraction), as a source for editing work, and as a motivator for clients to choose human translation due to bad AI experiences” (ELIS 2024: 24). Of course, the use of AI chatbots for translation work also comes with some important limitations that need our attention. One of them is that they need vast amounts of data to learn, meaning that their ability to translate depends largely on the amount of language data available to train on for any given language. Even though LLMs differ from all other machine learning-based MT systems in that they can be trained on monolingual data instead of bilingual corpora, thus increasing the amount of available data they can train on, they remain mostly trained on English data. Because of this, their translation performance is likely to be much less convincing for so-called low-resource languages compared with high-resource languages . In addition, AI chatbots are trained to identify and replicate language patterns. A natural corollary of this is that they lack “any form of genuine comprehension or the cognitive processes that humans use to understand language and context” (Moorkens et al. 2025: 189). Finally, when GenAI lacks the data it needs to produce part of a translation, it can “hallucinate”, i.e. make up content in the target language. These limitations notwithstanding, GenAI has started reshaping what it means to work as a professional translator in an industry that was already largely technology-driven, and where arguably automation has long been used as a way to increase the speed and reach of translation while reducing its cost. This has sometimes been to the detriment of working conditions and job satisfaction for human translators (see, for instance, Lambert & Walker 2024). This, in fact, makes it even more urgent for all of us translator educators to engage with GenAI so that we can, in the process, re-examine the role and agency of human translators within the translation process through that lens. In other words, GenAI should encourage us to rethink critically the value and values that humans bring to the translation process. As part of this, we must (re-)interrogate what we do on our translation programmes so that we empower the next generation of professional translators to achieve the kind of “human-centred augmented translation” that will benefit both individuals and society (O’Brien 2024: 391). Admittedly, however, for some of us translator educators the prospect of engaging with GenAI may initially have felt somewhat overwhelming. First of all, the timing may not have been ideal. OpenAI’s ChatGPT was released shortly after we came out blinking from a series of pandemic-related lockdowns, during iv Introduction which teaching happened mostly online. For some of us, this meant having to interact with (new) technologies that we may have felt were not always tailored to our own needs and/or that we weren’t always comfortable using. As a result, most of us got to experience first-hand the “technostress” many professional translators experience when they are asked to use translation tools and technologies they are not entirely comfortable with (Penet 2024). This is something we should take seriously, as technostress can “reduc[e] performance and har[m] individual wellbeing” (Koskinen 2020: 146). With its pedagogical approach, this edited volume therefore comes as an attempt to help alleviate feelings of stress among some of us as translator trainers. Ideally, we should all feel empowered to engage with GenAI on our programmes, where it also acts as a disruptor. Again, this was reflected in the findings of the latest ELIS reports. In its 2024 survey, university staff ranked GenAI implementation as the most widely shared challenge, with close to 90% of them seeing it as an issue. This prompted the report’s authors to comment that: “Generative AI and how to implement it in the universities’ programmes has taken the challenge chart of university staff by storm, topping even their concerns about the visibility of the profession and the eternal lack of time” (ELIS 2024: 27). If, the following year, visibility of the profession had reclaimed top spot in the list of challenges, it was still closely followed by GenAI implementation, which remained a concern for over 80% of university staff (ELIS 2025: 25). Yet, adapt and implement we must! This is because, still according to the report, in 2025 “[a]ctual MT use by language companies has increased […] and reaches now the magic mark of 50% of handled projects. AI makes its entry with an impressive 34%” (ELIS 2025: 35). In a similar vein, Slator, an online portal for language industry news and research, recently made the following point: ‘Skate to where the puck is going, not to where it is’ has always been good advice for ice hockey. But it is remarkably prescient for the language services industry as we prepare for the tsunami of disruption brought about by generative artificial intelligence (AI) and large language models, or LLMs. The speed of development has been frenetic since the release of OpenAI’s ChatGPT in November 2022. […] It would be a brave forecaster to predict how the corporate landscape will look five years from now—but operating from where the puck is today risks asking the wrong questions and missing the boat (Welocalize 2023). Even though this message was clearly intended for language service providers, it certainly holds some relevance for us educators. Whatever we may think or feel v Abbreviations Abbreviations 4EA Embodied, Embedded, Enacted, Extended, and Affective AI artificial intelligence ANT actor-network theory ASR automatic speech recognition AVT audiovisual translation CAT computer-aided translation CoT chain of thought (prompting) DDL data-driven learning DTM dynamic translation memory EFL English as a foreign language EMT European masters in translation ESL English as a second language FSL few-shot learning GBL game-based learning GenAI generative AI GPT generative pre-trained transformer GUMT guided use of MT HAI human-agent interaction HCAI human-centred AI HMI human-machine interaction ICL in context learning IT information technology L1 first language L2 second language LLM large language model LM language model LSC language service company LSP language service provider MKO more knowledgeable other MRU metacognitive resource use MT machine translation MTPE machine translation post editing NLP natural language processing NMT neural machine translation PE post editing PSF public service interpreting QC quality control QE quality estimation RLHF reinforcement learning using human feedback SDL self-directed learning SERF social ecology of responsibility framework SL source language ST source text STE simplified technical English TC translation competence T&I translation and interpreting TL target language TM translation memory TMS translation management system TS translation studies TT target text XAI explainable AI xiv Part I New paradigm: New skills & competences Chapter 1 Translation competence in the age of generative AI: Debates, dilemmas, directions Gary Masseya& Maureen Ehrensberger-Dowa aZHAW Zurich University of Applied Sciences, Switzerland (ret.) Until recently, the description and modelling of the competences and skills needed to translate successfully, and the ways in which they develop, have seen a steady evolution and predictable expansion, largely to accommodate technological advances and an increasing awareness of situatedness. However, the impact of neural MT and, now, generative AI (GenAI) has been unprecedented in rapidly transforming the core tasks of translators. Together with a proliferation of creative roles in a diversifying language industry, the volatility indicates a paradigm shift that is beginning to supplant even the once stable epithet “translator”. It also questions the efficacy of current descriptions of skills and confronts educators with dilemmas of balancing specialisation and generalisation, routinisation and adaptivity, core and transferable skills. This chapter considers relevant aspects of modelling competences and their development and examines related debates and dilemmas engaging educators and employers in the current and foreseen language-industry climate. It outlines directions for training in the age of translating with(out) GenAI, proposing an approach that, alongside core textual, interlingual translation and digital skills, combines transferable skills with human-machine/human-agent interaction (HMI/HAI) competence. Gary Massey & Maureen Ehrensberger-Dow. 2026. Translation competence in the age of generative AI: Debates, dilemmas, directions. In JC Penet, Joss Moorkens & Masaru Yamada (eds.), Teaching translation in the age of generative AI: New paradigm, new learning?, 3–26. Berlin: Language Science Press. DOI: 10.5281/zenodo.17641064 Gary Massey & Maureen Ehrensberger-Dow 1 Introduction 1.1 General What elements constitute translation competence (TC)1and how they can be developed have represented pivotal issues in applied Translation Studies since its inception (Holmes 1988: 77). They continue to trigger considerable debate amongst researchers, educators and practitioners, and they present educators and their organisations with dilemmas about how and where to target resources. In this chapter, we explore some debates and dilemmas triggered by changes to professional and educational practices brought about by generative artificial intelligence (GenAI). We also outline potential directions for human agents in the age of translating with and without GenAI. In our view, this necessitates an approach combining transferable skills with human-machine/human-agent interaction (HMI/HAI) competence, alongside “old-school” translation skills. Before proceeding to the debates, dilemmas and directions, however, we should define, describe and demarcate what exactly we mean by professional interlingual TC in the age of GenAI. 1.2 Definitions, descriptions and demarcation The word interlingual is used in Translation Studies and the language industry to designate a particular type of language mediation activity, namely that which takes place between natural languages and the particular cultures they represent. Alongside interpreting, translation represents a prototypical form of professional interlingual language mediation. In a language industry characterised by rapidly diversifying job titles and tasks (Bond 2018, Slator 2020: 11–17), the prototypical conceptualisation of what so-called (professional) translators do, i.e. translate content written in a source-language (SL) document into a target-language (TL) document, appears increasingly outmoded. Language professionals adaptively engage in a whole range of service provision as the lines between core and adjacent services blur (Angelone 2023, Angelone et al. 2024: 3–5). The major driver behind this state of affairs has been the technologisation of the industry as language service providers (LSPs) operate more and more downstream and upstream of previously core translation, localisation and interpret1This chapter applies the definitions of “skill” and “competence” from the latest EMT (2022: 3) competence framework. A skill is the “ability to apply knowledge and use know-how to complete tasks and solve problems”, while competence refers to “the proven ability to use knowledge, social and/or methodological abilities, in work or study situations and in professional and personal development”. 4 1 Translation competence in the age of generative AI ing activities. Indeed, the industry itself prefers the generic term linguist, and with good reason. Since 2016, the higher accuracy and fluency of neural machine translation (NMT) engines have led to more extensive automation of the translation process. Linguists have thus seen more deployment downstream of language transfer (or conversion) per se in post-editing (PE), output evaluation and quality assurance tasks. Most recently, the introduction of automated MT quality estimation (QE) to determine how much human PE is required is itself replacing more costly human MT quality evaluation (Slator 2024: 50–51). Upstream of transfer proper, linguists have been expanding their repertoire to perform tasks that include linguistic consultancy, MT pre-editing and multilingual content creation, for example in marketing, PR and corporate communications (often referred to as transcreation). Nevertheless, we shall retain the term translation, both in line with the title of the present volume and in recognition of the convincing argument made by do Carmo & Moorkens (2021) that the ultimate ethos of translation is to help communication flow between humans, regardless of the new technologies deployed and of the roles and activities these necessitate. GenAI has now entered the technological mix, and its applications clearly go far beyond interlingual mediation. Opportunities and risks have been identified in higher education in general (Atlas 2023, Gimpel et al. 2023, OECD 2023), where the potential of GenAI as an educational tool extends to teachers and students alike. Properly used, it can support course design, materials development, assessment, text and image production, coding, critical thinking and individualised learning. Improperly used, it can facilitate academic dishonesty, increase technological dependency, atrophy human skills and agency, undermine data protection and intellectual property rights, reinforce discrimination (due to data bias) and increase inequality (due to unequal access). Within the narrower confines of interlingual translation, GenAI differs decisively from AI-based language technology such as NMT in being capable of generating multilingual text, images and other content in response to prompts, including the prompt to translate. These prompts elicit statistically probable output generated by deep-learning models able to process and generate natural languages, known as large language models (LLMs). The prompt-response cycle can be iterative, allowing users to ask for more information or to provide more detailed prompts should the results require improvement (Gimpel et al. 2023: 22). Adequately structuring the instructions that prompt a GenAI model – also known as prompt design or, at the more technical level, prompt engineering – thus becomes central to its effective use. The skills and competences (see Footnote 1) necessary to embrace GenAI in the education and practice of professional 5 Gary Massey & Maureen Ehrensberger-Dow translators thus differ in major respects from those needed to master CAT tools and other translation technologies. It is indisputable that such skills and competences are essential to a language industry already transitioning from NMT to GenAI. Unbabel co-founder and CEO João Graça points out that LLMs are attracting more research and development, and can handle more complex tasks than NMT, such as automatic PE, source correction and cultural adaptation.2Lionbridge, one of the world’s leading LSPs, claims GenAI is “taking automated translation and localisation to new heights”.3Slator (2024: 52) reports that translation management systems (TMS) providers are already integrating GPT models to leverage LLMs’ capacity to rephrase, summarise and suggest new translations. LLMs can be used to incorporate QE into workflows and gauge PE effort, do PE or even provide linguistic insights throughout the translation workflow. Slator’s results also show two-thirds of MT providers offering fine-tuned LLMs and over 80% offering MT-LLM hybrid solutions. LSPs have the potential to become major providers and enablers of multilingual, multimodal GenAI content through content creation, validation, localisation, transmission and management. In keeping with the broadening portfolios of LSPs, multilingual text generation is currently “the most in-demand language AI application, after machine translation” (Slator 2024: 74). Fine-tuning and prompting LLMs to create and translate/localise content are emerging activities for language professionals to complement longer-established PE and quality assurance roles. Competent – or ideally expert – leverage of GenAI is a must for all those aiming to work in the language industry. But which skills and competences does and will this require? And in a dawning age of (language) work increasingly dominated by AI, what value can human intelligence and skills bring to bear? It is time to consider the debates around professional interlingual translation and the dilemmas that educators are facing. 2 Debates and dilemmas 2.1 Agency Asscher (2023) raised the issue of whether MT can actually be regarded as a proper point of interest in Translation Studies from a definitional perspective. He concludes that it is indeed compatible with both prescriptive, equivalenceoriented definitions of what translation is as well as descriptive definitions based 2Slatorpod #216, 12 July 2024 (https://www.youtube.com/watch?v=6smTEp3CXwQ) 3https://www.lionbridge.com/generative-ai/ 6 1 Translation competence in the age of generative AI largely on how translations are received. But he also points out that the perceived threat of AI to the social and professional status of translators, and of those who educate them (see also ELIS 2024: 24–25), may well redraw definitional boundaries and distinguish between MT and human translation on the basis of the perceived creative and moral authenticity of conscious agency (Asscher 2023: 14–16). Asscher’s argument draws on current positions on AI from the broader humanities and social sciences, such as moral philosophy (e.g. Sebastián & RudyHiller 2021). However, there is evidence of similar preoccupations in education policy and science, with emphasis falling on the importance of human agency, including teacher agency, and associated transferable skills – responsible action, critical thinking, systems thinking, logical reasoning, cultural agility, problem solving and emotion regulation (Gimpel et al. 2023, OECD 2023). Within the narrower confines of Translation Studies, it appears that agency is already emerging as the fulcrum around which the value of professional human translation can be measured. For example, Vieira (2020: 327–329) presents PE as a spectrum ranging from an MT-centred process typified by automatic PE to human-centred PE where humans have full control of output. Human agency moves back and forth along the spectrum as client, commission and employer demands require. The human-centred pole is more palatable to professional translators (Vieira 2020: 327), whose resistance to PE seems to derive from anxiety over a perceived loss of agency (Cadwell et al. 2018, Sakamoto 2019). In the discourse surrounding PE, agency has thus become a touchstone of professionalism and professional self-concept. Similarly, Rico & González Pastor (2022: 188) found in their pre-GenAI study of attitudes to teaching MT that the translation educators made much of the “human factor”. Though the data are unclear on precisely where they believe that agency resides, the participants place the translator at the centre of translation production. Interestingly, this prompts the researchers to posit that “MT has overcome its condition of tool”, and to claim that teaching MT should adopt a holistic approach “beyond an instrumentalist agenda that concentrates on the technical properties of the technologies” and which “evidences how the human factor is key to the translation process” (Rico & González Pastor 2022: 190, 193). Referring to other research, they also tentatively consider whether complementary effects of cognitive impairment or augmentation could influence competence development (Rico & González Pastor 2022: 192). One wider implication is that, even before the advent of GenAI, the paradigmatic human agency of the tool-user is being subverted by a more discerning and subtle appreciation of the rapidly evolving interactivity between translation technologies and their users. 7 Gary Massey & Maureen Ehrensberger-Dow offered for the translator’s approval or revision. Sometimes even perfect matches are not appropriate for the particular TT being produced (because of terminology, register, client style guidelines, etc.), so the translator has the choice of editing them or deleting them to translate from scratch. Many CAT tools now also have the option of accessing MT suggestions for segments or parts of segments that do not have good matches in the TM, either automatically or at the command of the translator.10 Usually much less typing is involved than in translating from scratch, but otherwise using a CAT tool to produce a TT requires a very similar set of skills. However, there are additional challenges in using a CAT tool to produce a TT. One is the focus on text segments (e.g. sentences or parts of sentences), which can make it more difficult for the translator to retain a sense of the whole text. Although CAT tools can provide cognitive relief by making it easier to be consistent, they can also increase the load by making it harder for the translator to maintain cohesion and coherence (Krüger 2016). Another challenge relates to the phenomenon of priming, in which “cognitive facilitation […] is triggered by linguistic overlap between earlier and current processing” (Vandepitte et al. 2018: 362). For example, suggestions that are not suitable can serve as inspirations for translators to produce versions that are (Farrell 2023). However, the sheer presence of a TM match or MT suggestion can also block cognitive processing and force the translator to exert extra effort to mitigate the effects of priming (Kolb 2024). 3.1.3 Using MT or GenAI for first drafts A translation process that includes producing a draft with an MT engine or GenAI is different from translating with a CAT tool or from scratch in certain ways, yet quite similar in others. The similarities lie in evaluating the translation solutions through a careful bilingual review and adapting the TT to be more appropriate. In all types of translation processes, thoroughly understanding the ST and recognising errors, misrepresentations and unmotivated omissions or additions in the draft are just as important as being able to edit the latter to produce a fluent, cohesive and coherent TT. Translation processes that incorporate GenAI shift the focus from writing to reading, prompting, evaluating and editing. Using a GenAI system with a simple prompt to translate the ST can be expected to produce a draft with features typical of MT output, the quality of which will depend on the particular engine being 10See Kappus (2024) for an overview of the development and convergence of translation technology. 14 1 Translation competence in the age of generative AI accessed. The process of improving such output through PE differs in many respects from editing or revision (do Carmo & Moorkens 2021). As well as a careful monolingual reading of the output, PE should always involve a bilingual review to eliminate any bias and inappropriate additions, omissions or “hallucinations” randomly introduced by the engine (see Dale et al. 2023). Working with GenAI in a more informed way necessitates a prompting process that provides detailed information about the purpose of the translation, client guidelines, target genre and audience. An iterative process of prompting, evaluation and reprompting to obtain suitable output could precede a prompting sequence to eliminate any problems such as terminological inconsistencies and unnecessary repetition. Similar issues to those mentioned above for MT output (i.e. bias, inappropriate additions and omissions, hallucinations) also need to be rectified. 3.2 Describing translation skills in the GenAI era It has been convincingly argued elsewhere (e.g. Nitzke & Hansen-Schirra 2021, O’Brien 2021) that using MT to produce high-quality TTs requires a special set of skills, and those arguments also hold when GenAI is simply used to produce a translation of the ST. However, informed use of GenAI includes the awareness that the pre-drafting and drafting phase of creating a TT requires well-considered prompting, and post-drafting demands careful editing. In the following, we consider textual and digital skills that should be foregrounded when training students to work with GenAI to produce TTs as well as the interlingual competence that needs to be developed for this. 3.2.1 Textual skills As Lafeber (2023: 43) has reported, professional translators should be able to “acquire subject-matter knowledge quickly […] understand complex topics, figure out obscure meaning, appreciate the authors’ intentions and the readers’ needs”. This high level of literacy suggests that a solid foundation in text analysis for translation (e.g. Nord 2005) is needed as well as the ability to recognise one’s own gaps in knowledge. Rather than hoping that an MT engine or GenAI will produce output that is easier to understand, students should be trained to identify difficult ST passages and to do the research needed to make sense of them. Gimpel et al. (2023: 37) point out that “[l]earners must possess adequate knowledge of the subject under scrutiny to achieve satisfactory outcomes”. In the context of translation, this means that students must have the skill to quickly search for 15 Gary Massey & Maureen Ehrensberger-Dow resources (e.g. reference words, parallel texts, videos, images) to help them understand the ST. They also need to have efficient, purposeful reading skills to evaluate the usefulness of the respective resource for their own comprehension and its potential as a GenAI prompt. Since GenAI can produce fluent, grammatically correct texts, the fostering of text production skills should focus on revising, editing and proofreading (see also Koponen et al. 2021). This represents a significant shift from the teaching of foreign languages and basic writing skills work that is still done in many entrylevel degree programmes, especially for translation into the L2 (Cerezo Herrero et al. 2021) and consistent with recent proposals for curriculum development (e.g. Sawyer et al. 2019). 3.2.2 Technological and digital skills The technology competence described in the EMT (2022: 9) framework potentially covers GenAI in stating that “students know how to use the most relevant IT applications […] and adapt rapidly to new tools and IT resources”. The need for a basic understanding of MT and data literacy are also referred to. However, GenAI only became widely available for translation after the latest iteration of the EMT framework appeared, so even the most recent descriptions seriously underspecify what type of knowledge is needed and which skills should be fostered to achieve competence when working with this tool. It is only since then that more detailed frameworks specifically targeting professional MT literacy, translation-oriented data literacy and AI literacy have begun to appear (e.g. Krüger 2024, Krüger & Hackenbuchner 2024), building on prior generic MT and data literacy frameworks (e.g. Bowker & Buitrago Ciro 2019). Especially relevant to the present context is the AI literacy framework for translation, interpreting, and specialised communication proposed by Krüger (2024). It contains various elements related to the functions and use of LLMs, some of which converge with the skills that we list in the final section. Indeed, scholars outside Translation Studies have already started arguing that all students need to acquire specific GenAI literacy. For example, Pretorius (2023) points out risks and challenges such as inaccurate results, unequal access to the technology, bias propagation and ethical concerns about sensitive data. We would add that students need to understand the consequences of the data sources used for the LLMs and the time lags associated with the latter’s knowledge updates, which can be in the order of months if not years. Perhaps the most noticeable change in technological and digital skills in an age of GenAI concerns the critical thinking required to formulate effective 16 1 Translation competence in the age of generative AI prompts, evaluate intermediate versions and adjust subsequent prompts accordingly. This also includes the ability to recognise the diminishing returns of continued prompting and the need to use one’s own textual skills to polish the TT. Working with GenAI is different from simply using language technology in that the intermediate versions can be considered a type of priming or mutual prompting between the conscious student/translator and the non-conscious AI agent. Only continuous critical reflection can counteract potentially overtrusting GenAI output during the translation process. Technological and digital translation skills in an age of GenAI might therefore be more broadly conceptualised as key components of an HMI/HAI competence capable of accounting for the particular reciprocities of human and AI agency. 3.2.3 Interlingual competence Prerequisites for professional translation have always included high reading proficiency in the SL as well as familiarity with the source culture and domain. No matter which aids are being used, the translator also needs high proficiency in the TL – both reading and revision skills – as well as in-depth knowledge of the domain and target culture. Interlingual transfer might be called for less often on the part of the professional translator deploying MT or GenAI, although they are always engaged in interlingual action. As Lafeber (2023: 43) found in her survey of institutional translators, professionals are expected to “achieve high levels of accuracy in their translations, conveying not only nuances but also intended effect […] draft well in their TL, compensating for poor wording in the original when appropriate while adhering to in-house conventions”. Just as actuaries attain high degrees of numeracy during their training but rarely do calculations by hand in their jobs, students should be given ample opportunity to translate with and without aids. The resulting interlingual competence in combination with HMI/HAI competence will contribute to their translation literacy (see also Massey 2021), effectively preparing them to work in the language industry in an age of GenAI. 3.3 Emerging roles in the language industry Translation training institutions are assumed to prepare students for the language industry, yet up to two-thirds of graduates may not actually work as translators (Hao & Pym 2023: 223). However, according to the same authors, many of them work in language-related professions. This suggests that programmes would do well to also prepare their students for jobs without “translator” in the 17 Gary Massey & Maureen Ehrensberger-Dow title. Even when translators are explicitly being recruited, though, a broad range of skills seems to be expected. In their analysis of job notices from 2005–2020, Prieto Ramos & Guzmán (2023: 53–55) found that the duties of translators at supranational and intergovernmental organisations also included (in order of overall average mention): assistance with other tasks, terminology work, revision, CAT-tool management and editing. On the topic of how automation is changing the translation profession, Pym & Torres-Simón (2021) discuss various recommendations for translators to focus on what machines cannot (yet) do, such as language service advice, service provision, language consulting and high-stakes communication. This is consistent with comments from industry observers that text production with GenAI is less suitable for “content that has a regulatory or technical purpose” (Slator 2024: 75). The same report also predicts increased demand for multilingual experts to revise corporate-generated output to meet expected quality levels (Slator 2024: 63). LSPs already offer a wide range of AI-related services (Slator 2024: 101), many of which require linguistic expertise that may still be difficult to find (cf. Faes et al. 2024). The latest European language industry survey (ELIS 2025: 34) shows the only increase in technology implementation over 2024 to be in GenAI. This represents an opportunity for both training institutions and their graduates and, encouragingly, 64% of the students participating in the same survey reported using GenAI in their training, 19% regularly (ELIS 2025: 37). This raises one last important question about those who train translation students. The teachers themselves patently require the (Gen)AI literacy needed to understand AI techniques, critically assess AI productions and recommendations, and use AI creatively in their teaching (OECD 2023: 401).11 And institutions need to have the staff development procedures in place that empower their teachers to do so. 4 Final take-aways Whether employed as translators, linguists, transcreators, localisers, consultants or language professionals, our graduates have an important role to play in the language industry despite or because of the introduction of GenAI. But they will need appropriate skills to do so. And those skills will, above all, have to accommodate the dynamic interactions of human and AI agency. Though degrees of domain specialisation are open to debate and will need to be determined by local 11Ideas for course design, materials and teaching scenarios that align well with the skills and competences outlined here can be found in Pym & Hao (2025). 18 1 Translation competence in the age of generative AI conditions, current TC models already adequately cover the old-school translation and textual skills that industry demands. But their transferable and technological and digital skills are underspecified, and the introduction of GenAI calls for greater precision. Krüger’s (2024) AI literacy framework is a response, outlining the digital competences needed to harness AI efficiently, ethically and sustainably. It covers technical foundations (operating principles, training, etc.), assessing AI’s usefulness for (domain-)specific tasks, interacting with (Gen)AI, implementing AI in workflows, and understanding ethical and societal aspects. Partially overlapping with elements of Krüger’s framework, our own brief recommendations represent a set of concrete priorities for embracing GenAI in translator education. We recommend that collaborative experiential learning should always integrate decisions about whether and how to deploy GenAI, together with critical evaluations of the resultant processes and products. Outcomes should focus on developing specific transferable skills: • critical thinking • adaptivity • creative problem-solving • cognitive and emotional self-regulation • self-efficacy • accountability and a sense of ethics • collaborative ability in human and AI interactions. The extension of the last point aims to develop what we have termed HMI/ HAI competence, which includes: • designing effective prompts (translation purpose, client specifications, target genre, audience, etc.), evaluating GenAI responses with supplementary research (where appropriate) and reprompting • recognising priming effects and mitigating negative ones • identifying and eliminating bias, additions, omissions, hallucinations • understanding LLM data sources and knowledge time-lags. None of the above should be regarded as optional add-ons but must receive at least equal weight to so-called core skills. Only then will our graduates be properly equipped for the GenAI age, whatever roles they take on. 19 Gary Massey & Maureen Ehrensberger-Dow References Angelone, Erik. 2023. Weaving adaptive expertise into translator training. In Gary Massey, Elsa Huertas-Barros & David Katan (eds.), The human translator in the 2020s, 60–73. London, UK: Routledge. Angelone, Erik, Gary Massey & Maureen Ehrensberger-Dow. 2024. Introduction: Contextualizing language industry studies. In Gary Massey, Maureen Ehrensberger-Dow & Erik Angelone (eds.), Handbook of the language industry: Contexts, resources and profiles, 1–13. Berlin, Germany: De Gruyter Mouton. Asscher, Omri. 2023. The position of machine translation in translation studies: A definitional perspective. Translation Spaces 12(1). 1–20. DOI: 10.1075/ts.22035. ass. Atlas, Stephan. 2023. ChatGPT for higher education and professional development: A guide to conversational AI. Tech. rep. Kingston: University of Rhode Island. https://digitalcommons.uri.edu/cba_facpubs/548. Bond, Esther. 2018. The stunning variety of job titles in the language industry. https: //slator.com/the-stunning-variety-of-job-titles-in-the-language-industry/. Bowker, Lynne & Jairo Buitrago Ciro. 2019. Machine translation and global research: Towards improved machine translation literacy in the scholarly community. Leeds, UK: Emerald Publishing. DOI: 10.1108/9781787567214. Bradshaw, Jeffrey M., Paul J. Feltovich & Matthew Johnson. 2011. Human-agent interaction. In Guy A. Boy (ed.), The handbook of human-machine interaction, 283–300. Boca Raton, USA: CRC Press. https : / / www . taylorfrancis . com / chapters/edit/10.1201/9781315557380-14/human%E2%80%93agent-interactionjeffrey-bradshaw-paul-feltovich-matthew-johnson. Brandtzaeg, Petter Bae, Marita Skjuve & Asbjørn Følstad. 2022. My AI friend: How users of a social chatbot understand their human–AI friendship. Human Communication Research 48. 404–429. DOI: 10.1093/hcr/hqac008. Cadwell, Patrick, Sharon O’Brien & Carlos S. C. Teixeira. 2018. Resistance and accommodation: Factors for the (non-) adoption of machine translation among professional translators. Perspectives 26(3). 301–321. DOI: 10.1080/0907676X. 2017.1337210. Cerezo Herrero, Enrique, Astrid Schmidhofer & Melita Koletnik. 2021. An LSP framework for translation and interpreting pedagogy. Revista de lenguas para fines especificos 37(2). 140–156. https://ojsspdc.ulpgc.es/ojs/index.php/LFE/ article/view/1406. Chaturvedi, Rkjul, Sanjeev Verma, Ronnie Das & Yogesh K. Dwivedi. 2023. Social companionship with artificial intelligence: Recent trends and future avenues. 20 1 Translation competence in the age of generative AI Technological Forecasting & Social Change 193. 122634. DOI: 10.1016/j.techfore. 2023.122634. Christensen, Tina Paulsen, Marian Flanagan & Anne Schjoldager. 2017. Mapping translation technology research in Translation Studies. An introduction to the thematic section. Hermes 56. 7–20. DOI: 10.7146/hjlcb.v0i56.97199. Dale, David, Elena Voita, Loic Barrault & Marta R. Costa-jussà. 2023. Detecting and mitigating hallucinations in machine translation: Model internal workings alone do well, sentence similarity even better. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, vol. 1, 36–50. https: //aclanthology.org/2023.acl-long.3.pdf. do Carmo, Félix & Joss Moorkens. 2021. Differentiating editing, post-editing and revision. In Maarit Koponen, Brian Mossop, Isabelle S. Robert & Giovanna Scocchera (eds.), Translation revision and post-editing: Industry practices and cognitive processes, 35–49. London, UK: Routledge. Eapen, Tojin T., Daniel J. Finkenstadt, Josh Folk & Lokesh Venkataswamy. 2023. How generative AI can augment human creativity. Harvard Business Review (July-August). https://hbr.org/2023/07/how-generative-ai-can-augmenthuman-creativity. Ehrensberger-Dow, Maureen, Andrea Hunziker Heeb, Gary Massey, Ursula Meidert, Silke Neumann & Heidrun Becker. 2016. An international survey of the ergonomics of professional translation. ILCEA 27. 1–23. http://ilcea.revues. org/4004. ELIS. 2024. ELIS 2024 European language industry survey. Tech. rep. ELIS Research. 1–57. https://elissurvey.org/wpcontent/uploads/2024/03/ELIS2024-Report.pdf. ELIS. 2025. ELIS 2025 European language industry survey. Tech. rep. ELIS Research. 1–53. http://elissurvey.org/wpcontent/uploads/2025/03/ELIS2025_Report.pdf. EMT. 2017. European master’s in translation competence framework 2017. Tech. rep. Brussels: European Commission. 1–12. EMT. 2022. European master’s in translation competence framework 2022. Tech. rep. Brussels: European Commission. 1–12. https://commission.europa.eu/ system/files/2022-11/emt_competence_fwk_2022_en.pdf. Esfandiari, Mohammad Reza, Nasrin Shokrpour & Forough Rahimi. 2019. An evaluation of the EMT: Compatibility with the professional translator’s needs. Cogent Arts & Humanities 6(1: 1601055). 1–17. DOI: 10 . 1080 / 23311983 . 2019 . 1601055. Faes, Florian, Alex Edwards & Anna Wyndham. 2024. Slator Q1 2024 research briefing. online. https://slator.com/event/slator-q1-2024-research-briefing/. 21 Gary Massey & Maureen Ehrensberger-Dow Faes, Florian & Gary Massey. 2024. Charting the language industry: Interview with an industry observer. In Handbook of the language industry, 17–31. Berlin, Germany: De Gruyter Mouton. DOI: 10.1515/9783110716047-002. Farrell, Michael. 2023. Do translators use machine translation and if so, how? Results of a survey held among professional translators. In Joss Moorkens & Vilelmini Sosoni (eds.), Proceedings of Translating and the Computer 44, 49–60. Geneva, Switzerland: Editions Tradulex. https://www.tradulex.com/varia/ TC44-luxembourg2022.pdf. Fügener, Andreas, Jörn Grahl, Alok Gupta & Wolfgang Ketter. 2021. Will humansin-the-loop become borgs? Merits and pitfalls of working with AI. MIS Quarterly 45(3). 1527–1556. DOI: 10.25300/MISQ/2021/16553. Gimpel, Henner, Caroline Ruiner, Manfred Schoch, Mareike Schoop, Luis Lämmermann, Nils Urbach, Kristina Hall, Torsten Eymann, Maximilian Röglinger, Steffen Vandirk, Alexander Mädche & Stefan Decker. 2023. Unlocking the power of generative AI models and systems such as GPT-4 and ChatGPT for higher education: A guide for students and lecturers. Working paper 02-2023. Hohenheim, Germany: University of Hohenheim. 4–46. https://hdl.handle.net/10419/ 270970. Guingrich, Rose E. & Michael S. A. Graziano. 2024. Chatbots as social companions: how people perceive consciousness, human likeness, and social health benefits in machines. https://arxiv.org/abs/2311.10599. Hao, Yu & Anthony Pym. 2023. Where do translation students go? A study of the employment and mobility of master graduates. The Interpreter and Translator Trainer 17(2). 211–229. DOI: 10.1080/1750399X.2022.2084595. Holmes, James S. 1988. Translated! Papers on literary translation and translation studies. Raymond van den Broeck (ed.), vol. 7 (Approaches to Translation Studies). Amsterdam, Netherlands: Rodopi. Hurtado Albir, Amparo (ed.). 2017. Researching translation competence by PACTE group. Amsterdam, Netherlands: John Benjamins. Kappus, Martin. 2024. Language technology developers. In Gary Massey, Maureen Ehrensberger-Dow & Erik Angelone (eds.), Handbook of the language industry: Contexts, resources and profiles, 121–141. Berlin, Germany: De Gruyter Mouton. Kiraly, Don. 2000. A social constructivist approach to translator education: Empowerment from theory to practice. London, UK: Routledge. Kiraly, Don. 2013. Towards a view of translator competence as an emergent phenomenon: Thinking outside the box(es) in translator education. In Don Kiraly, Silvia Hansen-Schirra & Karin Maksymski (eds.), New prospects and perspec22 1 Translation competence in the age of generative AI tives for educating language mediators, 197–224. Tübingen, Germany: Narr Verlag. Kolb, Waltraud. 2024. ‘I am a bit surprised’. Literary translation and post-editing processes compared. In Andrew Rothwell, Andy Way & Roy Youdale (eds.), Computer-assisted literary translation, 53–68. London, UK: Routledge. Koponen, Maarit, Brian Mossop, Isabelle S. Robert & Giovanna Scocchera. 2021. Translation revision and post-editing: Industry practices and cognitive processes. London, UK: Routledge. Krüger, Ralph. 2016. Contextualising computer-assisted translation tools and modelling their usability. Trans-kom 9(1). 114–148. https : / / d - nb . info / 1105491919/34. Krüger, Ralph. 2024. Outline of an artificial intelligence literacy framework for translation, interpreting and specialised communication. Lublin Studies in Modern Languages and Literature 48(3). 11–23. DOI: 10.17951/lsmll.2024.48.3.1123. Krüger, Ralph & Janiça Hackenbuchner. 2024. A competence matrix for machine translation-oriented data literacy teaching. Target 36(2). 245–275. DOI: 10.1075/ target.22127.kru. Lafeber, Anne. 2023. Skills and knowledge required of translators in institutional settings. In Tomáš Svoboda, Łucja Biel & Vilelmini Sosoni (eds.), Institutional translator training, 30–48. London, UK: Routledge. Latour, Bruno. 2005. Reassembling the social - An introduction to actor-networktheory. Oxford, UK: Oxford University Press. Massey, Gary. 2021. Translation literacy in additional language learning: Closing the conceptual divide between translation and language education. Bulletin suisse de linguistique appliquée 114. 31–45. DOI: 10.5169/seals-1030136. Mellinger, Christopher D. 2017. Translators and machine translation: Knowledge and skills gaps in translator pedagogy. The Interpreter and Translator Trainer 11(4). 280–293. DOI: 10.1080/1750399X.2017.1359760. Moorkens, Joss & Ana Guerberof Arenas. 2024. Artificial intelligence, automation and the language industry. In Gary Massey, Maureen Ehrensberger-Dow & Erik Angelone (eds.), Handbook of the language industry: Contexts, resources and profiles, 71–97. Berlin, Germany: De Gruyter Mouton. Nah, Fiona Fui-Hoon, Ruilin Zheng, Keng Siau & Langtao Chen. 2023. Generative AI and ChatGPT: Applications, challenges, and AI-human collaboration. Journal of Information Technology Case and Application Research 25(3). 277–304. DOI: 10.1080/15228053.2023.2233814. Nitzke, Jean & Silvia Hansen-Schirra. 2021. A short guide to post-editing. Berlin, Germany: Language Science Press. DOI: 10.5281/zenodo.5646896. 23 Erik Angelone Deliberate practice hinges on the learner’s intrinsic motivation, another core condition. Translators looking to advance along an expertise trajectory need to embrace working outside of their comfort zones and have a firm belief that deliberate practice will ultimately improve their performance. Deliberate practice takes immense time and effort. The oft-debated 10,000-hour ‘rule’ to becoming an ‘expert’ (Gladwell 2011) is often cited in attempts to quantify just how much time and effort are required. This would amount to “two and a half years of sustained effortful practice every day for five hours” (Shreve 2019: 173–174), raising questions regarding feasibility of this ‘rule’ and its place in the deliberate practice model. Deliberate practice calls for the learner’s commitment to “conscious performance monitoring” (Horn & Masunaga 2006: 601) and engaging in full concentration, as opposed to “mindless, routine performance” (Ericsson 2006: 692). This takes self-discipline and learner dedication to honing metacognitive capacities. Strategic scaffolding by the trainer, along with trainer/trainee dedication to the central ideas of cognitive constructivism (Piaget 1952), are instrumental in bolstering learner metacognition and performance monitoring processes. Figure 1 provides an overview of the core conditions of deliberate practice outlined in this chapter. It does not represent an exhaustive list of all conditions mentioned in the Expertise Studies literature, but rather focuses on those conditions taken up in Translation Studies to date. 3 Challenges to the implementation of deliberate practice It is worth noting here that empirical research on the benefits of deliberate practice on translation performance is still quite scant. This dearth can at least be partly explained by a series of inherent challenges associated with its implementation in formal training contexts. As mentioned in the previous section, several proponents of deliberate practice suggest that its benefits are best realised in contexts involving learners working one on one with individual trainers and in the absence of a pre-set curriculum (Ericsson et al. 1993: 367). This type of design does not readily align with the fashion in which translators are usually trained for a number of different reasons, starting with the financial constraint of needing to hire a personal trainer. Another constraint potentially standing in the way of meeting several of the core conditions of deliberate practice is the degree of learner heterogeneity commonly found in classroom-based training contexts. Students of translation often have widely varying levels of competence and experience, not to mention di30 2 Generative AI as a facilitator of deliberate practice in translator training immediate, informative feedback error correction opportunities "trainer" presence appropriate difficulty level intrinsic motivation conscious performance monitoring Core conditions of deliberate practice Figure 1: Core conditions of deliberate practice verse learning needs and interests. This can make it difficult for trainers to establish truly individualised learning objectives and corresponding tasks to meet them. The problem is exacerbated when student enrolments are high. Feedback, arguably the most important dimension of deliberate practice, often becomes less detailed and, out of necessity, much less immediate. Peer feedback and selffeedback activities can help address this gap, but immediacy, as a criterion for practice to be deliberate, often remains very difficult to obtain. Beyond time constraints, the deliberate practice condition of informative, immediate feedback might not be met if the kind of feedback being provided is not transferable in the sense of applying to future translation tasks. Oftentimes, assessment rubrics are used to mark up errors in accordance with various textual levels, such as grammar, word choice, and syntax. If the feedback given pertains to patterns along these lines that are applicable across tasks, it could be regarded as truly informative. If, on the other hand, feedback simply consists of marking up one-off errors pertaining to items that the translator may never encounter 31 Erik Angelone again in future translations, such as an isolated collocation error, it is relatively shallow and not particularly informative. Learner heterogeneity in large cohort settings also presents challenges when it comes to meeting the deliberate practice condition of making sure tasks are being undertaken at an appropriate level of difficulty. Translation practice courses are often informed by trainer intuition and prediction of appropriate difficulty level, in turn based on such facets as level of study, intake examination results, various cognitive process metrics (Sun & Shreve 2014), or previous student performance and corresponding ‘rich points’ (PACTE 2011). However, studies have shown a danger of misalignment between perceived or predicted problems and what actually proves to be problematic when translating (Angelone 2018). What is assumed to be difficult is actually not necessarily so, making attempts at predicting and setting an appropriate difficulty level challenging, particularly across a wide range of students with diverging needs. Furthermore, the idea of having students work outside of their comfort zone, or at the periphery of what they can realistically accomplish, for the purpose of advancing expertise may contradict approaches to translator training that embrace the predictability of staying within the learner’s comfort zone. According to deliberate practice guidelines, the learner’s intrinsic motivation needs to be constant. Having translation students work outside of their comfort zone in pedagogical contexts, particularly when grades are involved, runs the risk of hampering such motivation. The outcomes of deliberate practice would need to be gauged using metrics beyond formal grades, with a focus on helping learners become more self-reflective translators. Intrinsic motivation would come not so much from getting good grades or doing well in a given course, but rather from seeing the benefits of putting in hard work and embracing difficulty in order to become a better translator. 4 Generative AI as a facilitator of deliberate practice Against the backdrop of the conditions of deliberate practice presented in Section 2 and the constraints potentially standing in its way, as outlined in Section 3, the question remains: how can we best go about facilitating deliberate practice in translator training for purposes of expertise acquisition and advancement? In particular, how can we establish the requisite highly individualised, ‘coach’- oriented approach at the heart of deliberate practice? The pedagogical features of generative AI, in providing real-time feedback and prompt-driven interaction in a user-centred fashion, would seem to hold potential in this regard, as will be illustrated through a series of concrete scenarios in Section 6. 32 2 Generative AI as a facilitator of deliberate practice in translator training At the time of writing, we are still witnessing the dawn of generative AI as a vehicle for optimising translation, both in professional and pedagogical contexts. Its increasingly ubiquitous integration by LSP companies in project workflows has drawn attention to a need for artificial intelligence (AI) literacy (Krüger 2023), alongside MT literacy and data literacy in a broad sense. AI literacy can be defined as “a set of competencies that enables individuals to critically evaluate AI technologies, communicate and collaborate effectively with AI, and use AI as a tool online, at home, and in the workplace” (Long & Magerko 2020:2). In conjunction with discussions of translator performance, much of the discourse on AI intelligence to date has focused on how generative AI can be leveraged to facilitate human-in-the-loop translation, with an emphasis on the translation product. Outside of several of the chapters in this volume, relatively little discussion has been dedicated to the potential benefits of generative AI as a conversational agent, focusing less on generating translated content, and more so on enabling translators to reflect on their performance and engage in translation tasks to facilitate deliberate practice. Extending on sociocultural learning theories (Vygotsky 1965), some have come to regard a generative AI tool like ChatGPT as a ‘more knowledgeable other’ (MKO), in essence taking on the role of a personalised trainer that can “lead the learner from the zone of current development to the zone of proximal development – the space where one cannot quite master a content/task of their own, but they can with the help of an expert” (Stojanov 2023: 2). As an MKO, ChatGPT can address the aforementioned conditions of deliberate practice, particularly the presence of a one-on-one personalised trainer. Through real-time responses to learner prompts, ChatGPT ensures the immediacy of feedback that is so difficult to obtain in a larger enrolment classroom-based translator training environment. The learner’s intrinsic motivation is likely to be heightened when training is personalised and self-driven, rooted in immediate, informative feedback, and interactive, driven by their own prompts in relation to aspects of their own performance. 5 Generative AI as a scaffold for self-directed learning The utilisation of generative AI for purposes of deliberate practice calls for the learner to partake in self-directed learning (SDL), where “individuals take the initiative, with or without the help of others, in diagnosing their learning needs, setting learning goals, identifying resources, choosing appropriate learning strategies, and evaluating their learning outcomes” (Knowles 1975: 18). In this case, 33 Erik Angelone generative AI helps scaffold learning in line with learner prompts. Translation trainers well-versed in the conditions of deliberate practice can provide learners with valuable information on the nature of prompts they should enter. However, self-directed learners need to be “able, ready, and willing to prepare, execute and complete learning independently” (Jossberger et al. 2010: 419). The need for learner independence does not make the trainer superfluous, but rather shifts the focus of assignments undertaken and how they are assessed. Models of self-directed learning bear very close resemblance to the deliberate practice model. One such model that is widely cited in the literature consists of three closely interrelated dimensions: 1) self-management, 2) self-monitoring, and 3) motivation (Garrison 1997). Self-management involves the learner establishing concrete learning goals and managing learning resources to achieve these goals. In other words, they take control, deciding on the tasks in which they will engage. From a deliberate practice perspective, through strategic prompts, the translator can leverage generative AI to annotate errors in their translations. Generated annotations could then serve as a framework for the translator to selfdiscover the nature of the errors. Generative AI could then be prompted to provide similar translation tasks, with the goal of engaging the learner in deliberate practice centred around a certain error pattern (such as avoiding false cognates, erroneous literal translation, or problematic translationese at a syntactic level). Beyond error detection and mitigation, the translator can also use generative AI prompting to self-manage the difficulty level of the tasks they are undertaking. Section 6 provides more concrete scenarios and descriptions along these lines. Self-monitoring, the second component of Garrison’s model, pertains to the learner’s metacognitive processes. As an important dimension of self-directed learning, self-monitoring “requires learners to take responsibility to construct meanings” (Garrison 1997: 24). Through interactive feedback, generative AI can shed valuable light to help learners identify salient features of the translation task on which to focus their attention. For example, translators can enter a prompt asking generative AI to annotate source content that could be anticipated or predicted to present challenges in translation. Over the past decade, screen recording has found a place in process-oriented translator training for purposes of fostering self-monitoring and to enhance learner metacognition based on documentation of translation behaviors suggesting problems, including pausing, information retrieval, and revision (Angelone 2019). At present, generative AI tools like ChatGPT do not offer functionality where translators can upload screen recordings of their work for purposes of receiving analytic feedback at a granular level. Given recent advancements in this technology however, such as APIs that can provide 34 2 Generative AI as a facilitator of deliberate practice in translator training automated video summarisation, it is quite likely that utilisation of generative AI tools for such purposes is not too far away. The third component of Garrison’s self-directed learning model, motivation, directly aligns with motivation as a core condition of deliberate practice. Whereas deliberate practice regards motivation as the learner embracing challenge through an inherent desire to become better, Garrison draws attention to the importance of motivation for purposes of staying on task. The conversational interface of generative AI tools such as ChatGPT requires active participation on behalf of the learner, and, thereby, a heightened need to stay on task. However, without the physical presence of actual trainers or peers, staying on task, and motivation in general, is not a given. ChatGPT is not inclined to openly praise the performance of the learner to pique interest and motivation. Indeed, at present, the ChatGPT interface itself is quite basic, lacking any structural or discourse elements that might lend themselves well to inherently facilitating learner motivation. It will be interesting to see if this changes over time, perhaps in line with empirical user experience studies. 6 Application scenarios This chapter will now provide a series of concrete scenarios to illustrate how generative AI can be used to facilitate each of the core dimensions of deliberate practice put forward in Section 2 and to help translators advance along an expertise trajectory. The examples will be based on interaction with ChatGPT based on GPT-4o,1given the relative ubiquity and popularity of this particular generative AI tool at the time of writing. This focus on ChatGPT for illustrative purposes does not discredit the growing range of other generative AI tools available for use in a similar fashion. The translation scenario being used for purposes of contextualisation is the German-English translation of web content from a German private health insurance company,2translated for an international, English-speaking audience for informative purposes. The English translation was generated using DeepL.3 Figure 2 provides side-by-side alignment of the source and target content. As a point of departure, and to underscore the importance of strategic prompting when it comes to utilising GenAI for facilitating deliberate practice, (2) de1https://openai.com/index/gpt-4/ 2https://www.huk.de/gesundheit-vorsorge-vermoegen/krankenversicherung/ krankenvollversicherung.html#huk 3https://www.deepl.com/en/translator 35 Erik Angelone Figure 2: German-English translation of private health insurance content picts the relatively shallow output proposed by ChatGPT based on the following generic prompt: (1) “Set up translation training tasks with the goal of facilitating deliberate practice”. (2) ChatGPT output for setting up translation tasks to facilitate deliberate practice Setting up translation training tasks for deliberate practice involves focusing on specific skills, receiving immediate feedback, and gradually increasing difficulty. 1. Identify key skills and goals Skills: Vocabulary, grammar, idiomatic expressions, cultural context, speed, accuracy, style, and tone. Goals: Improve comprehension, enhance translation quality, increase speed without losing accuracy, and develop a nuanced understanding of both source and target languages. 2. Feedback and reflection Immediate Feedback: Provide prompt feedback on translations, highlighting errors and explaining corrections. Self-Assessment: Encourage learners to reflect on their work and identify areas for improvement. 3. Deliberate practice cycle Plan: Identify specific skills to focus on in each session. Practice: Perform targeted translation tasks. Feedback: Receive and review feedback from instructors or peers. 36 2 Generative AI as a facilitator of deliberate practice in translator training Reflect: Reflect on the feedback and identify areas for improvement. Repeat: Repeat the cycle with new tasks, gradually increasing difficulty. (3) depicts corresponding training tasks proposed by ChatGPT in parallel to the content in (2). (3) Week 1: Basic sentences Translate 10 simple sentences from the source language to the target language, with a focus on vocabulary and basic grammar. Week 2: Themed paragraphs Translate a paragraph related to a specific theme (e.g., technology), with a focus on thematic vocabulary and coherence. Week 3: Idiomatic expressions Translate sentences rich in idiomatic expressions, with a focus on cultural nuances and idiomatic accuracy. Week 4: Time translations Translate a 200-word article within 20 minutes, with a focus on speed and maintaining accuracy under time constraints. Week 5: Reverse translation Perform a back translation exercise, with a focus on understanding equivalence and identifying subtle changes in meaning. While tapping into the core conditions of deliberate practice put forward in the literature, this generic framework proposed by ChatGPT seems to focus first and foremost on group-based (rather than truly personalised) training and also seems to make the assumption that learners are starting ‘from scratch’ (moving from basic vocabulary to field-specific terminology, to cultural and idiomatic expressions, for example). Given the inherent learner heterogeneity described in Section 3, many translators would not directly benefit from such a lockstep, linear approach. As far as difficulty is concerned, it seems as if there is less focus on working at an appropriate difficulty level per se and more of an emphasis on incorporating a variety of potential sources of difficulty in a more fragmented sense (thematic vocabulary, idiomatic accuracy, time pressure, etc.). This is not to suggest that ChatGPT cannot personalise deliberate practice and cater to the needs of the individual user, but rather highlights the important, active role that the individual learner needs to play in order to benefit from generative AI. This 37 Erik Angelone starts with utilising more granular prompts in accordance with each specific condition of deliberate practice rather than taking a broad-brush approach such as that associated with the generic prompt that generated the output here. 7 Immediate, informative feedback At a very basic level, translators can utilise ChatGPT to get immediate, informative feedback on their performance by entering source content and their corresponding translations, along with correlating project specifications for the task (such as audience and purpose), in conjunction with the following prompt: (4) “Provide me with feedback on the following translation, German to English, of content from a private health insurance company in Germany for an international English-speaking audience, with the purpose of providing information”. (5) depicts the feedback obtained from ChatGPT in relation to this prompt. (5) ChatGPT feedback on a translation Your translation is clear and accurate, but I have a few suggestions to make it sound more natural for an international English-speaking audience. Here are the changes made: 1.Changed "advantages" to "benefits" for a more common term in health insurance. 2. Changed "tariffs" to "plans" to better fit the context. 3. Added a colon after "Your benefits at a glance" for better readability. 4. Capitalized "Private Health Insurer" for consistency with the title awarded. This feedback was generated in a matter of a few seconds, overcoming the time constraint often standing in the way of feedback being immediate in a typical translator training environment. As mentioned in Section 3, feedback becomes truly informative when it is applicable across translation tasks under similar conditions. With this in mind, the translator could enter the following prompt in response to the obtained feedback: (6) “Which of this feedback would be applicable when translating other private health insurance content from German into English in the future”? 38 2 Generative AI as a facilitator of deliberate practice in translator training (7) provides an overview of the response provided by ChatGPT, illustrating how feedback can become truly informative and transferable, as opposed to isolated to the one translation task at hand. (7) Informative, transferable feedback When translating content for other private health insurance companies in the future, you can apply the following feedback to ensure clarity, accuracy, and appropriateness for an international English-speaking audience: 1. Use Clear and Common Terms: Replace terms like "advantages" with "benefits" for more common understanding. 2. Ensure Readability and Natural Flow: Add colons or other punctuation to improve readability (e.g., "Your benefits at a glance:"). 3. Maintain Consistent Formatting: Use bullet points to list benefits for easy reading and better presentation. Interestingly, ChatGPT seems to attribute transferability to each piece of general feedback provided in response to the initial prompt, as depicted in (1). This suggests a potential tendency for ChatGPT to conflate task-specific and transferable feedback. However, this did not hold true in the context of a different simulated translation task, where some of the task-specific feedback was deliberately filtered out and not inherently regarded by ChatGPT as also being transferable. Nevertheless, from a prompting standpoint, for purposes of obtaining truly informative feedback, the translator may want to ask: “Which of this feedback is likely specific only to this one translation task and which of the feedback is transferable across translations in the same domain?”, with the latter being more beneficial from a deliberate practice perspective. 7.1 Error correction opportunities Immediate, informative feedback, as a core condition of deliberate practice, provides the translator with granular insight into the nature of errors. In hopes of advancing along an expertise trajectory, it is just as important for the translator to have opportunities to correct errors they have made by virtue of having them annotated but not spelled out, and to translate additional source content in which 39 Erik Angelone Ericsson, K. Anders & Andreas C. Lehmann. 1996. Expert and exceptional performance: Evidence of maximal adaptation to task constraints. Annual Review of Psychology 47. 273–305. DOI: 10.1146/annurev.psych.47.1.273. Garrison, D. R. 1997. Self-directed learning: Toward a comprehensive model. Adult Education Quarterly 48(1). 18–33. DOI: 10.1177/074171369704800103. Gladwell, Malcolm. 2011. Outliers: The story of success. New York, USA: Back Bay Books. Hatano, Giyoo & Kayako Inagaki. 1986. Two courses of expertise. In Harold Stevenson, Hiroshi Azuma & Kenji Hakuta (eds.), Child development and education in Japan, 262–272. New York, USA: W. Y. Freeman & Co. Horn, John & Hiromi Masunaga. 2006. A merging theory of expertise and intelligence. In The Cambridge handbook of expertise and expert performance, 587–611. Cambridge, UK: Cambridge University Press. DOI: 10.1017/CBO9780511816796. 034. Jossberger, Helen, Saskia Brand-Gruwel, Henny Boshuizen & Margje van de Wiel. 2010. The challenge of self-directed and self-regulated learning in vocational education: A theoretical analysis and synthesis of requirements. Journal of Vocational Education & Training 62(4). 415–440. DOI: 10.1080/13636820.2010. 523479. Knowles, Malcolm. 1975. Self-directed learning: A guide for learners and teachers. Chicago, USA: Follett. Krüger, Ralph. 2023. Artificial intelligence literacy for the language industry – with particular emphasis on recent large language models such as GPT-4. Lebende Sprachen 68(2). 283–330. DOI: 10.1515/les-2023-0024. Long, Duri & Brian Magerko. 2020. What is AI literacy? Competencies and design considerations. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20), 1–16. Honolulu, USA: Association for Computing Machinery. DOI: 10.1145/3313831.3376727. https://doi.org/10.1145/3313831. 3376727. Miller, Scott D., Daryl Chow, Bruce E. Wampold, Mark A. Hubble, A. C. Del Re, Cynthia Maeschalck & Susanne Bargmann. 2020. To be or not to be (an expert)? Revisiting the role of deliberate practice in improving performance. High Ability Studies 31(1). 5–15. DOI: 10.1080/13598139.2018.1519410. https: //doi.org/10.1080/13598139.2018.1519410. PACTE. 2011. Results of the validation of the PACTE translation competence model: Translation problems and translation competence. In Cecilia Alvstad, Adelina Hild & Elisabet Tiselius (eds.), Methods and strategies of process research: Integrative approaches to translation studies, 317–343. Amsterdam, Netherlands: John Benjamins. 46 2 Generative AI as a facilitator of deliberate practice in translator training Piaget, Jean. 1952. The origins of intelligence in children. New York, USA: International Universities Press. Shreve, Gregory. 2006. The deliberate practice: Translation and expertise. Journal of Translation Studies 9(1). 27–42. Shreve, Gregory. 2018. Levels of explanation and translation expertise. Hermes 57. 97–108. Shreve, Gregory. 2019. Professional translator development from an expertise perspective. In Erik Angelone, Maureen Ehrensberger-Dow & Gary Massey (eds.), The Bloomsbury companion to language industry studies. London, UK: Bloomsbury Academic. Stojanov, Ana. 2023. Learning with ChatGPT 3.5 as a more knowledgeable other: An autoethnographic study. International Journal of Educational Technology in Higher Education 20(1). 35. DOI: 10.1186/s41239-023-00404-7. Sun, Sanjun & Gregory Shreve. 2014. Measuring translation difficulty: An empirical study. Target 26(1). 98–127. Vygotsky, Lev. 1965. Thought and language. Cambridge, USA: MIT Press. 47 Chapter 3 AI Literacy: The concept of suitability and core translation skills Ramon Inglada Heriot-Watt University, United Kingdom The release of ChatGPT in November 2022 reignited the debate on the future of the translation profession and, consequently, on how translator training programmes should change and adapt. Similar to discussions that occurred with the introduction of neural machine translation a few years earlier, some argued that core translation skills, such as language knowledge, translation ability and cultural expertise that have long been essential components of many translator training programmes, had now been rendered obsolete by the arrival of generative AI (GenAI). Inspired by the concept of machine translation literacy, a case will be made that these core skills (together with some other complementary skills, such as selection and assessment) are absolutely essential in order to ascertain whether the content produced by GenAI tools is not simply accurate or inaccurate, but, more importantly, suitable for the requirements of any given translation project, as specified in the relevant translation brief. Examples will be given of specific AI-based tasks (such as translation of content, terminology extraction, multilingual glossary creation and machine translation postediting) in which the suitability of the results cannot be determined without recourse to so-called traditional core translation skills. 1 Machine translation literacy and artificial intelligence literacy The concept of machine translation (MT) literacy, introduced by Bowker & Buitrago Ciro (2019), has attracted a lot of interest, both in academia and in Ramon Inglada. 2026. AI Literacy: The concept of suitability and core translation skills. In JC Penet, Joss Moorkens & Masaru Yamada (eds.), Teaching translation in the age of generative AI: New paradigm, new learning?, 49–64. Berlin: Language Science Press. DOI: 10.5281/zenodo.17641068 Ramon Inglada the language industry, as well as in other sectors. Furthermore, Bowker’s MT Literacy project1aims to educate users on the dos and don’ts of using MT output and its infographics have been very popular on social media. It also highlights the importance of confidentiality and privacy when using programs such as DeepL and Google Translate, with an emphasis on different use cases of free online MT systems. The project’s focus on enhancing digital literacy skills and the responsible, ethical, and sustainable use of MT systems can be used as the basis to further develop and adapt the concept of Artificial Intelligence (AI) literacy. Long & Magerko (2020: 2) defined AI literacy as “a set of competencies that enables individuals to critically evaluate AI technologies; communicate and collaborate effectively with AI; and use AI as a tool online, at home, and in the workplace”. In this chapter, this definition will be used as the basis for a different definition specifically created with the translator training context in mind. This approach aligns with (Krüger 2024) framework for AI literacy in translation, which emphasises the need for translators to develop competencies in understanding and effectively using AI technologies in their work. It is undeniable that the release of ChatGPT in November 2022 generated a lot of hype. It also reignited the debate on the future of the translation profession and, consequently, on how translator training programmes should change in order to adapt to this latest technological advancement. Similar to discussions that occurred with the introduction of neural machine translation (NMT) systems a few years earlier, some argued that core translation skills that have long been essential components of translator training programmes had now been rendered obsolete by the arrival of GenAI. However, it could be argued that this is an oversimplistic approach. This chapter advocates for technological advancement, as it can bring many benefits to society and offer solutions to some of the greatest challenges humanity currently faces. In the field of translation, when used sensibly as part of welldesigned workflows, computer-assisted translation (CAT) tools and MT can be a great asset for professional translators. The same applies to GenAI chatbots based on large language models (LLMs), such as OpenAI’s ChatGPT, Microsoft Copilot and Google Gemini. However, given how recent this technology is and the pace at which it evolves, it could be argued that the translation industry might struggle to agree on what a ‘sensible’ and ‘well-designed’ AI-based workflow constitutes in the field of professional translation. In any case, it is always worth keeping in mind that no technology-based solution is infallible. This obviously applies to ChatGPT (and to the other LLMs) and it could be said that, to 1https://sites.google.com/view/machinetranslationliteracy/ 50 3 AI Literacy: The concept of suitability and core translation skills a certain extent, even ChatGPT itself reminds us to be vigilant and use our own judgment and common sense. At the time of writing, this is the message that appears underneath ChatGPT’s chat interface: ChatGPT can make mistakes. Consider checking important information. 2 Core translation skills The purpose of this chapter is not to provide an exhaustive list of core translation skills, describing what they are and why they can be considered as key skills all translators should have, as many others have already done this very effectively in the past. Instead, it will focus on three very general skills which, arguably, most translation scholars would generally agree are essential for all professional translators (and which, therefore, should feature prominently in all translator training programmes). This is also done for the sake of simplicity as, ultimately, these skills are mere examples. This selection of core skills has been based on the reference standards for translator training set out in the European Master’s in Translation (EMT) Competence Framework (2022) and on the translation competence model created by PACTE (2003). Different translator training programmes, institutions and settings could naturally select other skills that they would consider as essential. These chosen three core skills are: • Linguistic knowledge • Translation ability • Cultural expertise Core translation skills such as the ones mentioned above should not be allowed to disappear from translator training programmes, but rather should now be considered more important than ever, precisely because of the advent of GenAI. This is not to say that translator training curricula should not be adapted: they should, particularly as concepts such as augmented translation and human in the loop are becoming increasingly important, and discussions around translators now becoming ‘language specialists’ or ‘language experts’ are happening with increasing frequency. However, the debate should not be about these core skills becoming obsolete, but about how essential they are when GenAI is used in translation. 51 Ramon Inglada 3 The concept of suitability over correctness When translators or translation students use ChatGPT or other similar tools to assist them in their translation tasks, the focus should not be on whether the AIgenerated output is ‘correct’ (something that could also entail a degree of subjectivity), but rather on whether such output is suitable (or fit-for-purpose) for any given set of translation requirements, as described in the translation brief, for instance in terms of tone, register, target audience, vocabulary and text type. The priority should be placed on the concept of ‘suitability’ of AI-generated content, rather than on its ‘correctness’. This cannot be successfully achieved if the aforementioned core skills (as a minimum, linguistic knowledge, translation ability and cultural expertise) are not present and well honed. Translators would struggle to decide whether AI-generated content is suitable for any given translation project (and its set of unique needs) if they do not know how to translate (and they have not developed their core translation skills). Using GenAI chatbots is easy; using their output critically requires thought. 4 New and complementary skills It could also be argued that not only are core translation skills still essential, but that they should also be complemented with the ‘new’ skills of selection and assessment. These are obviously not new skills — after all, many translators have been already selecting and assessing translation memory matches for decades — but they have now become fundamental when using GenAI in translation. Professional and trainee translators need to be able to select among the alternatives offered by ChatGPT and other similar tools. They also need to be able to critically assess the suitability of the output being presented to them. Therefore, a renewed emphasis should be given to selection and assessment as core skills to be integrated into and developed in translator training programmes. The reasons why ‘traditional’ core language, translation and cultural expertise skills should still be a pivotal part of translator training programmes have already been discussed, in combination with the reasons why they should also be complemented by the not-so-new skills of selection and assessment. However, is there any additional new skill that should be added to this list of core skills for the training of translators in the new era of GenAI? The answer is a resounding yes. This skill is prompting (also known as prompt engineering). Prompt engineering can be defined as the process of designing, crafting, and refining inputs to elicit specific responses from a GenAI model, aiming to optimize interaction outcomes through careful consideration of the prompts (Bozkurt 2024). The model 52 3 AI Literacy: The concept of suitability and core translation skills then generates a response based on the input it receives. Effective prompting is crucial for obtaining relevant, coherent and suitable answers. Translator training programmes should seek to integrate prompting skills into their curricula, so that future translators can ensure a greater degree of suitability in the output generated by GenAI. When adapting existing translator training programmes (or when designing new offerings), an emphasis should be placed on processes aimed at the careful creation of effective prompts (or sequences of prompts) with different roles/personalities, different levels of complexity or even different degrees of creativity. 5 AI Literacy in translation: a simple definition Based on all the principles discussed before, AI Literacy for Translation could be defined as the combined application of a definite set of basic core skills in order to maximise the usefulness and relevance of GenAI output and ensure its suitability in relation to the requirements set out by the translation brief in any given translation task. The basic core skills are linguistic knowledge, translation ability and cultural expertise, complemented by selection and assessment. All these skills should be applied to GenAI output produced as a result of effective prompting. The overarching notion that underpins the application of AI Literacy in translation is that of suitability, rather than ‘correctness’, of the generated output. 6 AI Literacy applied to practical language and translation tasks Five examples will now be provided of translation or translation-related tasks, chosen to represent some of the tasks that professional translators carry out on a regular basis (and which, as such, are also commonly observed within the confines of the translation classroom). The concept of AI Literacy will be applied to these examples. This means that, in all cases, an argument will be presented to emphasise the importance of relying on the three generic core skills of linguistic knowledge, translation ability and cultural expertise (presented earlier in this chapter) in order to select the most appropriate GenAI output and assess the suitability of this output. An example of the prompt (or prompts) used to request the completion of the task in hand by the GenAI chatbot (ChatGPT based on GPT 3.5) will also be provided for each example. 53 Ramon Inglada In the first example, ChatGPT itself was used to request the creation of a sentence related to renewableenergy containing some grammar issues and awkward word groupings. In the other four examples, the three initial paragraphs from the English version of the Wikipedia article on the topic of renewable energy (from July 2024) were used as the reference source text. These five examples are: sentence reformulation, translation of content, terminology extraction, multilingual glossary creation and MT post-editing (MTPE). 6.1 Example 1: Sentence reformulation For our first example, I will start with one of the simplest tasks LLMs can be used for. This task has also been chosen as an introductory example as anecdotal evidence seems to suggest that this represents one of the primary ways in which GenAI chatbots are commonly utilised, especially by non-native speakers of the language they are engaging with. ChatGPT can be a valuable tool for reformulating badly written sentences. By providing a poorly constructed sentence to ChatGPT, users can receive suggested revisions that address spelling mistakes, grammar errors, improve clarity and enhance overall readability. ChatGPT leverages its model of language patterns and grammar rules to generate alternative phrasing that is more coherent and idiomatic. Users can input sentences from various fields, such as renewable energy in our example, to receive tailored suggestions for improvement. With ChatGPT’s assistance, individuals can refine their writing skills, produce clearer communication and convey their messages effectively across diverse contexts. In this example, the following prompt was used to ask ChatGPT to create the sentence I would use as a basis for reformulation (see (1)): (1) Prompt used to elicit a poorly constructed sentence from ChatGPT 3.5. (Created by Ramon Inglada) This was the sentence resulting from the prompt above: (2) ‘The sun is giving us muchly energy, so we should put many solar panels to catch it all’. The following prompt in (3) was then used to request an improved version of the sentence, correcting any grammar issues and changing any unnatural word collocations: 54 3 AI Literacy: The concept of suitability and core translation skills (3) Sentence reformulation prompt in ChatGPT. (Created by Ramon Inglada) This was ChatGPT’s reply to our request: (4) ‘The sun provides us with abundant energy, so we should install numerous solar panels to capture it efficiently’. It could certainly be argued that this version corrects existing grammar issues and uses more idiomatic collocations to convey the intended meaning more clearly. However, it would only be possible to do that if our knowledge of the English language is advanced enough. If I now try to apply the concept of AI Literacy in this first example, it is immediately evident that I would need to rely on one of the aforementioned core skills (in this case, knowledge of the English language) firstly to realise that the original sentence contains some issues, and secondly to assess the suitability of the suggested ‘improved’ version provided by the LLM. Furthermore, in this specific example, ChatGPT decided that three elements in the original sentence needed to be changed. These are ‘muchly energy’, ‘many solar panels’ and ‘catch it all’. If my English language skills were insufficient, I could simply accept all three suggested improvements, assuming (or even hoping) that the resulting sentence is now much more suitable for my needs. However, I could also decide to select only some of these changes, so I would only keep those that I (and not ChatGPT) consider as suitable. I could even accept them all and then go on to further modify them to end up with a final sentence which would be the collaborative result between me and the LLM. This process exemplifies the concept of “centaur tasks” as described by Mollick (2023), where there is a clear division of labour between human and AI, leveraging the strengths of each. Once more, the decision on whether this final sentence is the most suitable option for our needs is something that can only be achieved if our core skills have been developed enough. 6.2 Example 2: Translation of content This second example is intended to cover the use of LLMs as MT tools. While LLMs were not in principle designed to be used as MT providers, this is certainly one of the many tasks these tools can perform. It is also probably one of the first uses that comes to mind when one thinks about the potential uses of GenAI in 55 Ramon Inglada attempt to overcome these high levels of hype. This objective is critical because it helps to ensure that the expectations of AI are realistic. It could be said that, in some cases, the hype surrounding AI has led to unrealistic expectations. Yes, GenAI can be very useful in the translation field and it has many potentially useful applications (as we have seen in this chapter and as many others have explained elsewhere). However, it is still essential for translation students to acquire core translation skills and for professional translators to continue developing and enhancing them throughout their careers. By applying the principles of AI Literacy, it is hoped that we can contribute to ensuring that expectations linked to the use of GenAI in translation are realistic and that the technology is used in a way that is beneficial to all stakeholders (including translators). It is equally important to note that AI is not a panacea for all translation ‘problems’, and it is essential to use it in conjunction with other technologies and approaches. GenAI should become another tool in the extensive translator’s toolbox, but it should not aim at replacing translators altogether. It is also important to mention that the concept of AI Literacy and its key principles have been adapted with flexibility firmly in mind. They are not necessarily meant to be taken literally as they have been presented in this chapter (although it is hoped that this would indeed be possible), but rather as guiding principles open to interpretation and adaptation based on specific contexts and evolving requirements and perspectives. As Long & Magerko (2020) discuss, AI Literacy encompasses a range of competencies that can be tailored to different needs and contexts. Similarly, Krüger (2024) outlines an AI Literacy framework that emphasises the adaptability of several skills. The core skills (linguistic knowledge, translation ability, and cultural expertise), complementary skills (selection and assessment), and the new skill (prompting) that have been presented in this chapter as underpinning the concept of AI Literacy, have been included as examples. The number of skills is flexible and can be changed based on requirements. In addition, these skills can be tailored or modified to align with diverse needs. AI Literacy is not meant to be an unmovable or fixed entity, but rather a dynamic and evolving concept that adapts to changing circumstances, contexts, and priorities. A point that warrants reiterated emphasis is that GenAI should not be viewed as a replacement for human translators. Instead, it should be considered as a tool that could be used to augment the work of translators. The concept of AI Literacy and its underlying principles emphasise the need to use core translation skills when using GenAI for translation tasks. These skills include not only linguistic knowledge but also the ability to understand the context, culture, and nuances of the language. Translators can add value by leveraging these skills to improve the 62 3 AI Literacy: The concept of suitability and core translation skills quality and suitability of the output produced by GenAI. Moreover, the potential for AI to hallucinate should never be underestimated. Waldo & Boussard (2024) analysed the reasons why LLMs hallucinate and concluded that these models encountered challenges with topics for which limited data was available online, often generating inaccurate responses presented in a realistic format without acknowledging the inaccuracies. The responsibility to find issues and fix them, and to make the output produced by GenAI suitable for any set of specific needs at any given time, is routinely placed on translators. Without the skills the conform the basis of AI Literacy, this would be an impossible endeavour. Debates around the emergence, evolution and implementation of AI in our modern societies should be based on the overarching principles of ethics, sustainability and data privacy. Discussions around AI Literacy should not be an exception and, as the application of GenAI in translation continues to evolve, attention should be devoted to concerns about its ethical implications, environmental impact, and potential impact on the translation profession. The development and deployment of AI in the field of translation raise a wide range of ethical concerns, including those related to accuracy, bias, transparency, authorship, privacy and personal data. GenAI in translation has many potentially useful applications and this has been widely discussed in this chapter. However, its ethical, sustainability, and professional implications must be carefully considered and addressed to ensure that AI-based translation-related tasks and workflows are developed and used responsibly and sustainably. It is hoped that promoting the notion of AI Literacy (and integrating this concept in translator training curricula) will also contribute to better decision-making and to more mature debate (and scrutiny) among all stakeholders in the translation industry. As mentioned earlier in this chapter, using GenAI chatbots is easy, but using their output critically requires thought. Ultimately, AI Literacy in the field of translation (and, more specifically, in the field of translation training) aims at equipping future translators —and other language professionals— with this fundamental ability for critical thinking. References Bowker, Lynne & Jairo Buitrago Ciro. 2019. Machine translation and global research: Towards improved machine translation literacy in the scholarly community. Leeds, UK: Emerald Publishing. DOI: 10.1108/9781787567214. Bozkurt, Aras. 2024. Tell me your prompts and I will make them true: The alchemy of prompt engineering and generative AI. Open Praxis 16(2). 111–118. DOI: 10.55982/openpraxis.16.2.661. 63 Ramon Inglada EMT. 2022. European master’s in translation competence framework 2022. Tech. rep. Brussels: European Commission. 1–12. https://commission.europa.eu/ system/files/2022-11/emt_competence_fwk_2022_en.pdf. Krüger, Ralph. 2024. Outline of an artificial intelligence literacy framework for translation, interpreting and specialised communication. Lublin Studies in Modern Languages and Literature 48(3). 11–23. DOI: 10.17951/lsmll.2024.48.3.1123. Long, Duri & Brian Magerko. 2020. What is AI literacy? Competencies and design considerations. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20), 1–16. Honolulu, USA: Association for Computing Machinery. DOI: 10.1145/3313831.3376727. https://doi.org/10.1145/3313831. 3376727. Mollick, Ethan. 2023. On-boarding your AI Intern. (25 November, 2024). https: //www.oneusefulthing.org/p/on-boarding-your-ai-intern. PACTE. 2003. Building a translation competence model. In Fabio Alves (ed.), Triangulating translation: Perspectives in process oriented research (45), 43–66. Amsterdam, Netherlands: John Benjamins. DOI: 10.1075/btl.45.06pac. Waldo, Jim & Soline Boussard. 2024. GPTs and hallucination: Why do large language models hallucinate? Queue 22(4). 19–33. DOI: 10.1145/3688007. Yamada, Masaru. 2023. Optimizing machine translation through prompt engineering: An investigation into ChatGPT’s customizability. In Masaru Yamada & Félix do Carmo (eds.), Proceedings of Machine Translation Summit XIX, Vol. 2: Users Track, 195–204. Macau, China: Asia-Pacific Association for Machine Translation. https://aclanthology.org/2023.mtsummit-users.19. 64 Part II New paradigm: New knowledge Chapter 4 Teaching translation students about data in the age of generative AI Lynne Bowker Université Laval, Canada AI translation tools are now a key part of translation education, but many educators are searching for effective ways to teach the essentials of these tools to students with no background in computer science. This chapter explains why corpora make a good entry point to learning about AI translation tools, and it explores how science communication techniques such as framing, analogies and visualization can be used to help translation educators and students come to grips with data and machine learning. 1 Introduction Though translator education often takes place in an arts or humanities faculty, technology has been a vital part of the translation profession, and hence of translator education programmes, for at least thirty years. Over this period, translators and translation students have shown remarkable resilience as they adapt to new tools and new technology-based ways of working. However, this does not mean that teaching technologies to translation students is easy, and educating the educators can be a particular challenge (Bowker 2023, Kenny 2020). As the pace of new tool releases gets faster, it can be difficult for translator educators to know where to begin. This chapter proposes that teaching translation students about data is one of the key building blocks in preparing them to use translation technology effectively. In the case of translation, data frequently takes the form of texts organised into corpora. This chapter therefore begins with a brief review of corpora and corpus-based tools, noting how AI translation tools have influenced Lynne Bowker. 2026. Teaching translation students about data in the age of generative AI. in JC Penet, Joss Moorkens & Masaru Yamada (eds.), Teaching translation in the age of generative AI: New paradigm, new learning?, 67–85. Berlin: Language Science Press. DOI: 10.5281/zenodo.17641070 Lynne Bowker the nature of corpora. Next, the chapter outlines why a science communication approach, rather than traditional scientific communication, could be useful in the context of translation technology education. Three science communication techniques – framing, analogies and visualisation – are combined with examples to demonstrate how science communication could be adapted to teach translation students about data-related topics in the age of AI. 2 Corpora and corpus-based translation tools 2.1 The changing characteristics of corpora in the age of AI Over the past few decades, translators have seen the introduction and integration of a range of different tools, including concordancers (Zanettin 2023), term extractors (Korkontzelos & Ananiadou 2022), translation memory systems (Melby & Wright 2023), machine translation systems (Way 2020), and most recently, generative AI (GenAI) systems (Siu 2024). The various tools have become increasingly sophisticated with regard to their capabilities and their underlying architecture, but a common feature for all of the tools named above is that their core functionality revolves around processing data in the form of texts. Collections of texts are usually referred to as corpora, and these can take different forms depending on the nature of the texts and the way that they are organised (McEnery 2022). Translation tools that process corpora are frequently described as corpus-based or data-driven tools (Wang et al. 2022). As pointed out by Isabelle et al. (1993: 205), “existing translations contain more solutions to more translation problems than any other available resource”. Therefore, one type of corpus that has been used very often by translation tools is the bilingual parallel corpus (Simard 2020). In this type of corpus, a collection of source texts are aligned – usually at sentence level – with their counterpart target texts. In other words, each sentence in the source text is linked to its corresponding translation in the target text. Some translation tools might also use monolingual corpora of original texts in the source and/or target language to act as a linguistic model for that language. For students who have already learned about concordancers, term extractors or translation memory systems, the notion of a corpus is already familiar because the corpus is a very visible resource in such tools. When using these tools, the translator or translation student often has a hand in creating the corpus or may need to upload the corpus that has been provided by the client or educator. Owing to their familiarity, corpora make a good starting point for learning about neural machine translation or GenAI tools. 68 4 Teaching translation students about data in the age of generative AI According to Bowker & Pearson (2002: 9), A corpus can be described as a large collection of authentic texts that have been gathered in electronic form according to a specific set of criteria. There are four important characteristics to note here: ‘authentic’, ‘electronic’, ‘large’ and ‘specific criteria’. While the general notion of a corpus remained relatively stable in the period before AI translation tools appeared, the introduction of these tools has ushered in some changes with regard to the features of corpora. Therefore, it is important for translation students to understand how AI tools have influenced and altered the nature of corpora. 2.2 Machine-readable form The need for the corpus to be in electronic or machine-readable form has not changed. Indeed, we could say it is more important than ever since AI translation tools take on an even greater degree of corpus processing than do tools such as concordancers or translation memory systems. While concordancers and translation memory systems conduct pattern matching and then sort and display information for the tool users to interpret, AI translation tools go further by attempting to interpret the results and present fully formed translation solutions. 2.3 Size Corpora are used to reveal linguistic patterns, which only become apparent when there are multiple examples of a given linguistic phenomenon. Therefore, another feature of corpora is that they are usually very large collections of text. However, our understanding of what constitutes “large” has evolved over time. The first generation of corpora created in the 1960s contained hundreds of thousands of words and were mainly consulted by linguists (McEnery 2022). These linguists used corpus analysis tools (e.g. concordancers) to help them sort and display the text data, but the linguists were still responsible for interpreting it. Today, the corpora used to power AI tools contain hundreds of billions of words (Hughes 2023). This is in large part because these AI tools do not understand text in the way that people do, and they need a much larger number of examples in order to predict patterns with confidence. However, this need for extremely large corpora is influencing other characteristics of corpora, such as the texts selected for inclusion (see Section 2.4) and sometimes even the authenticity of the texts (see Section 2.5). 69 Lynne Bowker The availability of machine-readable texts can differ dramatically from one language to another, creating disparities with regard to an AI tool’s performance in different languages. A high-resource situation occurs when it is relatively straightforward to gather a large number of high-quality resources for a given language or language pair. For instance, both English and French are widely used languages, and there is a lot of translation activity between these two languages. As a result, it is not too difficult to compile monolingual and bilingual parallel corpora for these languages, and they are thus referred to as high-resource languages. In contrast, a low-resource situation can occur when languages (or language varieties) are less widely used, such as some of the Indigenous languages of the Americas or the variety of French used in Canada. For languages or language varieties of limited diffusion, it can be more challenging to build a large corpus. Moreover, even if two languages have a large number of speakers (e.g. Russian and Hindi), there may not be a lot of translation activity between them, making it hard to create bilingual parallel corpora for this language pair. Therefore, languages, language varieties or language pairs for which there are few corpora available are described as being low resource. 2.4 Specific criteria As emphasised by McEnery (2022), in order to be most useful, a corpus cannot consist of texts that have been gathered at random or in a purely opportunistic way. Rather, the texts in a corpus are selected because they correspond to specific criteria and are representative of a larger set of texts with those characteristics. One clear criterion in the context of translation is that the texts should be of high quality. Beyond this, there are many different options for designing a corpus depending on its intended purpose, but the key point here is that the choice of which texts to include is motivated. For example, in the context of a corpus to be used for translation, it could be important to select texts that are on a given topic, of a certain text type or register, or from a particular time period. Evidence of the importance of corpus design can be seen in the way that translators construct translation memory databases, such as by creating different databases (or adding relevant metadata) for different domains or for different clients (e.g. to respect their preferred terminology or house style). In this way, they can restrict a search to texts that have specific features (Melby & Wright 2023). Likewise, neural machine translation tools are known to achieve better quality when the corpus is adapted for a specific domain (Chu & Wang 2018). However, as noted in section Section 2.3, AI translation tools need to have an enormous number of texts in the corpus. As a result, it can be challenging to 70 4 Teaching translation students about data in the age of generative AI achieve the necessary size while trying to be selective about the content. The low-resource situation described in section Section 2.3 can be further complicated when it comes to finding certain text types or texts on very specialised topics in less widely used or translated languages and language varieties. A consequence of not having a large enough high-quality corpus is that the AI tool does not have enough good examples to draw on, and so the tool’s performance may be poorer in low-resource situations (Way 2025). If lower quality texts are included in the corpus in order to increase the size, then the tool may generate low-quality translations. The implications of insufficient quantity and quality of texts in training corpora used by AI translation tools is discussed in more detail in sections Section 4.2 and Section 4.3. 2.5 Authenticity Finally, the need to have authentic texts in the corpus used to be sacrosanct. For instance, McEnery (2022) describes a corpus as “a large body of linguistic evidence composed of attested language use” (494) and “a collection of naturally occurring language data” (495). In the case of translation, there is a desire to have high-quality data, which means using texts that have been translated by language professionals. As observed by Kenny (2011: 2), the reason that the developers of translation tools use corpora of human translations to train their systems is because such corpora are assumed to contain good answers to translation problems; and they are assumed to contain good answers precisely because they contain translations performed by human beings. One consequence of the need for bigger and bigger bilingual parallel corpora for training AI translation tools has been that such authentic high-quality human translated texts have become increasingly valuable commodities. This in turn has raised many ethical questions about ownership of translation data and permission to use it, prompting an explicit need to discuss these issues with translation students. Moorkens (2022) contains a detailed examination of such ethical issues, along with suggestions for how these can be integrated into translator education (see Section 4.1). Another response to the potential shortage of text needed for constructing very large corpora for use with AI translation tools has been to set aside the longestablished tradition of using authentic data and to explore the use of synthetic data. In the context of translation, synthetic data is created by using a machine translation tool to translate additional texts, and then adding these machinetranslated texts to the corpus (Sennrich et al. 2016). The quality of synthetic data can vary, and while the need for additional data tends to be for low-resource 71 Lynne Bowker To see examples of images of bananas, translation educators and students can navigate to open access image sites such as Pixabay.com or Unsplash.com and use the search term “bananas”. To perform well, an AI tool would need to be provided with many, many examples of pictures of bananas in order to be able to generate a new image based on these examples. People typically need fewer examples from which to learn. A fun exercise to do in class is to ask the students to look at a dozen pictures of bananas on an open access site, and then to imagine that they are an AI image generation tool that has been asked to generate (i.e., draw) their own image in response to the following prompt: “Draw a banana.” This will not likely pose too much of a challenge for the students – even those who are not artistically inclined! As a follow-up, share with students the results of the Daniel Hook (2023) experiments in which he asked the GenAI tool Midjourney to generate images of a banana. What Hook found was that the AI tool only generated images with two or more bananas, rather than a single banana. Hook refined the prompt several times asking for “a banana”, “a single banana”, “one banana”, and so on, but the tool continued to generate images containing multiple bananas. This example, which Hook (2023) dubs “The lone banana problem”, can be used as a simple yet effective way of explaining the issue of data bias. If the training data does not contain a sufficient number of images of individual bananas, then the AI tool will not be able to generate an image of a single banana. Of course, the problem can be fixed by adding more data (i.e., images of lone bananas), but as Hook warns, we do not always realise what gaps are there. Hook (2023) also stresses that it is important to recognise that AI tools do not understand the world the way that people do: they only understand commonly occurring patterns. While humans are amazing pattern matchers, that skill is augmented by common sense (in many but not all cases), context and an evolved and subtle understanding of the physical world around us. AIs don’t yet have those augmentations – they are pure pattern matching power. And hence, they are only as good as the data that we input into the training set and hence can be no more than the statistical average of those inputs. In the lone banana problem, the statistics suggested that bananas only appear in twos (or more) and so the AI could not imagine a single banana. (Hook 2023, n.p.) 4.3.3 Example 2: The snow detector As a follow-up to the lone banana problem, it may be useful to share the example of an AI tool whose task is image classification. This example comes from Ribeiro 78 4 Teaching translation students about data in the age of generative AI et al. (2016), who trained an AI tool using two different sets of images – ten images of husky dogs and ten images of wolves. After the training session, the researchers tested the AI tool’s ability to classify images correctly by showing it ten new images (five huskies and five wolves) that had not been part of the initial training data. The AI tool classified most of the images correctly, but it made two errors. In investigating the cause of the errors, the researchers found that the AI tool was making decisions not based on the animals in the pictures but based on the background. In the training data, all the pictures of the wolves contained snow, but none of the pictures of huskies contained snow. Therefore, in the test phase, if the image showed an animal against a snowy background, the AI tool classified the image as a wolf, but if there was no snow, then the AI tool classified the image as a husky. In other words, the AI tool was looking for patterns, but the pattern that it used as a basis for its decision-making was not a sensible one. The snow detector is a funny example, but the medical literature contains more serious examples of similar phenomena. In Narla et al. (2018), researchers tried to train an AI tool to distinguish between images of cancerous and noncancerous skin lesions. However, in the training corpus, the majority of images of cancerous lesions also contained a ruler to measure the size of the lesion. During the testing phase, it became clear that the AI tool was equating the absence of a ruler with the absence of cancer. In another example, an AI system was trained to distinguish lung X-rays of people with pneumonia from lung X-rays of people with COVID-19 (Roberts et al. 2021). Unfortunately, the training data containing examples of pneumonia consisted mainly of pediatric patients, while the training data with examples of COVID-19 was from mainly adult patients. As a result, the AI tool learned to distinguish between children and adults rather than between pneumonia and COVID-19. 4.3.4 Connecting the dots: From images to translation Though humorous, the lone banana and snow detector examples are intended to drive home the very serious point that AI tools do not think, understand or even process information the way that people do. By focusing on relatively straightforward, concrete and visual tasks such as image generation and image classification, these examples are easy for students to relate to, and they can then be connected to translation-related issues. For example, the lone banana problem can lead into a discussion of data bias and of what happens when a translation corpus has gaps in the training data. For instance, Vanmassenhove et al. (2018) have demonstrated how data bias can lead to issues of gender bias in translation, 79 Lynne Bowker while Bowker & Blain (2022) point out how a lack of data in one language variety can lead to content from another variety being used erroneously in a translation. In addition to reinforcing the lesson about data bias, the snow detector example, along with the more serious examples of the ruler detector and the child detector, can be used to promote discussions on opacity and what can go wrong when it is not clear what patterns the AI tool is using as the basis for decisions. These examples can also feed discussions about risk assessment – an important element of any decision about whether or not to use an AI translation tool (Koponen & Nurminen 2024). In some cases, the translation task at hand might be relatively low-stakes (e.g. translating a text for the purposes of entertainment), and so using an AI tool could be a good option. However, in other circumstances, the translation task might be one with higher stakes, where the consequences of a poor translation could negatively affect someone’s health or life (Way 2013). Another type of information that can be important as part of risk assessment is understanding whether the situation is more likely to be a high-resource or a low-resource situation, given that the performance of the AI tool is likely to be better for high-resource languages, language varieties, language pairs and domains, and poorer in low-resource situations (see Section 2.3). 5 Conclusion AI translation tools are rapidly inserting themselves into the translation industry and must also be addressed in translator education. The technology behind these tools is very sophisticated and may be intimidating for translation educators and students alike, given that most do not have a background in computer science. However, most language professionals do not need to understand all the details of how AI translation tools work in order to use them in an informed and responsible way. In contrast, it is extremely useful for language professionals to understand how these tools interact with data, and how the data used for machine learning can affect tool performance and translation quality. In the context of AI translation tools, the data is typically a corpus. Translation students are already familiar with corpora, which are also used with other types of translation tools (e.g. concordancers, term extractors, translation memory systems). However, students need to learn how AI tools are impacting the characteristics of corpora. Rather than approaching the teaching of AI translation tools from the more traditional perspective of scientific or expert-to-expert communication, translation educators may find that science communication techniques (e.g. framing, 80 4 Teaching translation students about data in the age of generative AI analogies, visualisation) are more useful for teaching students about data and machine learning. In the spirit of transversal learning, translation educators may even find that developing effective techniques to teach their translation students about data and its role in AI translation tools can pay off in other ways. As Borowiec (2023: 1) laments, Despite there being many excellent reasons for scientists to engage in science communication, they often lack the tools to do so. […] Scientists frequently cite lack of training and/or confidence in their science communication skills as a barrier to their participation in public-facing activities. Since many translator educators are also researchers, they may be able to apply these science communication techniques to their research-oriented activities. Moreover, Ehrensberger-Dow et al. (2023) suggest that translator trainers are well placed to take on a role such as machine translation literacy consultant, which could involve helping clients (who are unlikely to be computer scientists) to understand how AI tools process data or how data resources can be curated to improve the tools’ performance. References Borowiec, Brittney G. 2023. Ten simple rules for scientists engaging in science communication. PLoS Computational Biology 19(7). e1011251. DOI: 10 . 1371 / journal.pcbi.1011251. Bowker, Lynne. 2023. Translation technologies. In Chan Sin-wai (ed.), Routledge encyclopedia of translation technology, 2nd edn., 95–113. London, UK: Routledge. Bowker, Lynne & Frédéric Blain. 2022. When French becomes Canadian French: The curious case of localizing COVID-19 terms with Microsoft Translator. The Journal of Internationalization and Localization 9(1). 1–37. DOI: 10.1075/jial. 22007.bow. Bowker, Lynne & Jennifer Pearson. 2002. Working with specialized language: A practical guide to using corpora. London, UK: Routledge. DOI: 10 . 4324 / 9780203469255. Briva-Iglesias, Vicent & Sharon O’Brien. 2022. The language engineer: A transversal, emerging role for the automation age. Quaderns de Filologia: Estudis Lingüístics XXVII. 17–48. 81 Lynne Bowker Burns, T. W., D. J. O’Connor & S. M. Stocklmayer. 2003. Science communication: A contemporary definition. Public Understanding of Science 12(2). 183–202. DOI: 10.1177/09636625030122004. Chu, Chenhui & Rui Wang. 2018. A survey of domain adaptation for neural machine translation. In Proceedings of the 27th international conference on computational linguistics, 1304–1319. Santa Fe, USA: Association for Computational Linguistics. Cooke, Steven J., Austin J. Gallagher, Natalie M. Sopinka, Vivian M. Nguyen, Rachel A. Skubel, Neil Hammerschlag, Sarah Boon, Nathan Young & Andy J. Danylchuk. 2017. Considerations for effective science communication. Facets 2. 233–248. DOI: 10.1139/facets-2016-0055. Ehrensberger-Dow, Maureen, Alice Delorme Benites & Caroline Lehr. 2023. A new role for translators and trainers: MT literacy consultants. The Interpreter and Translator Trainer 17(3). 393–411. DOI: 10.1080/1750399X.2023.2237328. Flach, Peter. 2012. Machine learning. Cambridge, UK: Cambridge University Press. Hackenbuchner, Janiça & Ralph Krüger. 2023. DataLitMT – teaching data literacy in the context of machine translation literacy. In Mary Nurminen, Judith Brenner, Maarit Koponen, Sirku Latomaa, Mikhail Mikhailov, Frederike Schierl, Tharindu Ranasinghe, Eva Vanmassenhove, Sergi Alvarez Vidal, Nora Aranberri, Mara Nunziatini, Carla Parra Escartín, Mikel Forcada, Maja Popovic, Carolina Scarton & Helena Moníz (eds.), Proceedings of the 24th annual conference of the European Association for Machine Translation, 285–293. Tampere, Finland: European Association for Machine Translation. Hanauska, Monika. 2019. Historical aspects of external science communication. In Annette Leßmöllmann, Marcelo Dascal & Thomas Gloning (eds.), Science communication, 585–600. Berlin, Germany: De Gruyter Mouton. DOI: 10.1515/ 9783110255522-027. Henville, Letitia. 2020. Effective science communication deconstructed. University Affairs. https://universityaffairs.ca/career-advice/ask-dr-editor/effectivescience-communication-deconstructed/. Hook, Daniel. 2023. The lone banana problem. Digital Science. https:/ /www. digitalscience.com/tldr/article/thelonebananaproblemorthenewprogramming-speaking-ai/. Hughes, Alex. 2023. ChatGPT: everything you need to know about OpenAI’s GPT-4 tool. https://www.sciencefocus.com/future-technology/gpt-3. Isabelle, Pierre, Marc Dymetman, George Foster, Jean-Marc Jutras, Elliott Macklovitch, François Perrault, Xiaobo Ren & Michel Simard. 1993. Translation analysis and translation automation. In Proceedings of the fifth interna82 4 Teaching translation students about data in the age of generative AI tional conference on theoretical and methodological issues in machine translation of natural languages: Association for computational linguistics, 201–217. Kenny, Dorothy. 2007. Translation memories and parallel corpora: Challenges for the translation trainer. In Dorothy Kenny & Kyongjoo Ryou (eds.), Across boundaries: International perspectives on translation studies, 192–208. Newcastle-Upon-Tyne, UK: Cambridge Scholars Publishing. Kenny, Dorothy. 2011. The ethics of machine translation. In Sybille Ferner (ed.), Proceedings of the 20th New Zealand Society of Translators and Interpreters national conference. Auckland, New Zealand: New Zealand Society of Translators & Interpreters. Kenny, Dorothy. 2018. Machine translation. In Piers Rawling & Philip Wilson (eds.), The Routledge handbook of translation and philosophy, 428–445. London, UK: Routledge. Kenny, Dorothy. 2020. Technology and translator training. In Minako O’Hagan (ed.), The Routledge handbook of translation and technology, 498–515. London, UK: Routledge. Koponen, Maarit & Mary Nurminen. 2024. Risk management for content delivery via raw machine translation. In Marion Winters, Sharon Deane-Cox & Ursula Böser (eds.), Translation, interpretation and technological change, 111–135. London, UK: Bloomsbury. Korkontzelos, Ioannis & Sophia Ananiadou. 2022. Term extraction. In Ruslan Mitkov (ed.), The Oxford handbook of computational linguistics, 2nd edn., 991– 1012. Oxford, UK: Oxford University Press. Lankford, Séamus, Haithem Afli & Andy Way. 2023. adaptNMT: An open-source, language-agnostic development environment for neural machine translation. Language Resources and Evaluation 57. 1671–1696. DOI: 10.1007/s10579-02309671-2. Marshman, Elizabeth & Lynne Bowker. 2012. Translation technologies as seen through the eyes of educators and students: Harmonizing views with the help of a centralized teaching and learning resource. In Séverine HubscherDavidson & Michał Borodo (eds.), Global trends in translator and interpreter training, 69–95. London, UK: Continuum. McEnery, Tony. 2022. Corpora. In Ruslan Mitkov (ed.), The Oxford handbook of computational linguistics, 2nd edn., 494–507. Oxford, UK: Oxford University Press. Melby, Alan K. & Sue Ellen Wright. 2023. Translation memory. In Chan Sinwai (ed.), Routledge encyclopedia of translation technology, 2nd edn., 766–781. London, UK: Routledge. 83 Lynne Bowker Moorkens, Joss. 2022. Ethics and machine translation. In Dorothy Kenny (ed.), Machine translation for everyone, 121–140. Berlin, Germany: Language Science Press. Narla, Akhila, Brett Kuprel, Kavita Sarin, Roberto Novoa & Justin Ko. 2018. Automated classification of skin lesions: From pixels to practice. Journal of Investigative Dermatology 138(10). 2108–2110. DOI: 10.1016/j.jid.2018.06.175. National Academies of Sciences, Engineering, and Medicine. 2017. Communicating science effectively: A research agenda. Washington DC, USA: The National Academies Press. DOI: 10.17226/23674. https://nap.nationalacademies.org/ catalog/23674/communicating-science-effectively-a-research-agenda. Ribeiro, Marco T., Sameer Singh & Carlos Guestrin. 2016. Why should I trust you? Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: Association for computing machinery, 1135–1144. New York, USA. DOI: 10.1145/ 2939672.2939778. Ridley, Michael. 2025. Human-centered explainable artificial intelligence: An annual review of information science and technology (ARIST) paper. Journal of the Association for Information Science and Technology 76(1). 98–120. DOI: 10.1002/asi.24889. Roberts, Michael, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I. Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, Jonathan R. Weir-McCall, Zhongzhao Teng, AixCovnet Effrossyni Gkrania-Klostas, James H. F. Rudd, Evis Sala & CarolaBibiane Schönlieb. 2021. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nature Machine Intelligence 3. 199–217. DOI: 10.1038/ s42256-021-00307-0. Sennrich, Rico, Barry Haddow & Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th annual meeting of the Association for Computational Linguistics (volume 1): Association for computational linguistics, 86–96. Simard, Michel. 2020. Building and using parallel text for translation. In Minako O’Hagan (ed.), The Routledge handbook of translation and technology, 78–90. London, UK: Routledge. Siu, Sai Cheong. 2024. Revolutionising translation with AI: Unravelling neural machine translation and generative pre-trained large language models. In Yuhong Peng, Huihui Huang & Defeng Li (eds.), New advances in translation technology, 29–54. Singapore: Springer. DOI: 10.1007/978-981-97-2958-6_3. 84 4 Teaching translation students about data in the age of generative AI Vanmassenhove, Eva, Christian Hardmeier & Andy Way. 2018. Getting gender right in neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 3003–3008. Brussels, Belgium: Association for Computational Linguistics. Wang, Haifeng, Hua Wu, Zhongjun He, Liang Huang & Kenneth W. Church. 2022. Progress in machine translation. Engineering 18. 143–153. DOI: 10.1016/j.eng. 2021.03.023. Way, Andy. 2013. Traditional and emerging use-cases for machine translation. In Proceedings of Translating and the Computer 35. London, UK: ASLIB. https: //aclanthology.org/2013.tc-1.12/. Way, Andy. 2020. Machine translation: Where are we at today? In Erik Angelone, Maureen Ehrensberger-Dow & Gary Massey (eds.), The Bloomsbury companion to language industry studies, 311–332. London, UK: Bloomsbury. Way, Andy. 2025. What does the future hold for translation technologies? In Stefan Baumgarten & Michael Tieber (eds.), The Routledge handbook of translation technology and society, 448–461. London, UK: Routledge. Zanettin, Federico. 2023. Concordancing. In Chan Sin-wai (ed.), Routledge encyclopedia of translation technology, 2nd edn., 498–511. London, UK: Routledge. 85 Chapter 5 Teaching translation with AI: Bridging theory and practice through prompt engineering Masaru Yamada Rikkyo University, Japan This chapter explores the innovative application of large language models (LLMs) in translator training, focusing on the use of few-shot prompts and chain-ofthought prompting. It proposes a novel approach that integrates metalanguages and concepts of Translation Studies into prompt engineering, moving beyond traditional natural language processing goals of improving machine translation quality. The chapter demonstrates how this method can create interactive and engaging learning experiences for translation students, allowing them to explore various translation strategies and develop critical thinking skills. Through concrete examples, the chapter illustrates the potential of LLMs to generate diverse translation variations and provide insightful analyses of translation processes. While acknowledging limitations and the need for critical evaluation, the research emphasises the positive and proactive possibilities of LLMs in translator training. This approach not only bridges the gap between translation theory and practice but also opens new avenues for autonomous learning and the development of essential skills for future translators in the AI era. 1 Introduction Recent advancements in artificial intelligence (AI), particularly the emergence of Large Language Models (LLMs), have generated extensive debate regarding their benefits and drawbacks. In response, the UK’s Institute of Translation and Interpreting (ITI) has articulated the Slow Translation Manifesto (ITI 2024). Drawing Masaru Yamada. 2026. Teaching translation with AI: Bridging theory and practice through prompt engineering. In JC Penet, Joss Moorkens & Masaru Yamada (eds.), Teaching translation in the age of generative AI: New paradigm, new learning?, 87–104. Berlin: Language Science Press. DOI: 10.5281/zenodo.17641072 Masaru Yamada treme and may lack realism, it can be likened to the specific style and voice used in a particular company’s user manuals in a real-world context. The purpose of this approach is twofold: to observe how the LLM adapts its output based on the provided translation memory, and to encourage learners to critically analyse how they can learn from and apply specific linguistic features, tones, and styles in their own translation work. This exercise, while using an exaggerated example, aims to heighten awareness of the importance of adapting language to specific contexts and audiences in translation practice. This prepared sample set was then given as a few-shot prompt to the LLM, instructing it to translate a new source text following the corpus. The LLM used was Claude 3.5 Sonnet. To provide a comparison, the new source text was also translated using the MT service DeepL. Subsequently, the LLM translations and the MT translations were compared against the original corpus. Additionally, evaluations were conducted using the automatic evaluation tool COMET2The results of these comparisons and evaluations are presented below. LLM Prompt 5.1: Below are provided the English text [source text], in which an elderly individual living in contemporary America reflects on their life, and its Japanese translation [target text]. Based on this corpus, please translate the [new source text] into Japanese. [source text] “When I was young, life was so different. We didn’t have all these gadgets and technology. We had to work hard for everything we had. But, those were good times too. I remember walking miles to school, playing outside until dusk, and everyone knew each other in the community.” [target text] 「若い頃はのう、生活はまったく違っとったんじゃ。今みたい なガジェットやテクノロジーはなかったんじゃ。何でも手に入れるために一生 懸命働かなければならんかったんじゃ。でも、それもまた良い時代だったんじ ゃよ。学校まで何マイルも歩いて、夕暮れまで外で遊んで、みんながコミュニ ティの中でお互いを知っとったんじゃ。」 2We used WMT20-COMET for this evaluation. 94 5 Teaching translation with AI [new source text] “I miss those simple days. We didn’t have much, but we were happy. I spent my summers helping my father on the farm and learning about life from him. Sundays were always special, with family gatherings and big dinners. As I grew older, the world changed rapidly, and sometimes it’s hard to keep up. But I cherish those memories, they keep me grounded.” Figure 1: Claude vs. DeepL. According to the COMET scores in Figure 1, Claude’s scores range from 0.360 to 0.995, while DeepL’s scores range from 0.633 to 1.051. These scores indicate generally high-quality translations, as scores closer to 1.0 typically reflect strong alignment with the reference text. On average, DeepL achieves higher scores at 0.626 compared to Claude at 0.499. However, when evaluated by a human translator, a stark stylistic difference becomes evident between the two. Claude Translation skilfully captures the distinctive elderly speech patterns found in the reference translations. Using a small corpus, it creates the impression that the same elderly person is continuing the conversation seamlessly. In contrast, DeepL Translation employs a completely standard Japanese tone, making it appear as though a different person is speaking, thereby disrupting the flow of the monologue. This discrepancy highlights a limitation of the COMET scoring system: it fails to account for stylistic features or cultural nuances such as those present in Japanese “elderly speech patterns”. 95 Masaru Yamada In an educational context, it is crucial for human translators to first anticipate the linguistic features that can be learned from the corpus provided in the prompt, and then compare these expectations with the LLM’s response. Additionally, comparing with MT allows reflection on how these predictions differ from actual outputs. When COMET scores are available, it may be beneficial to critically analyse whether these scores accurately reflect our predictions. As mentioned above, in this case, it appeared that COMET did not adequately reflect the situation, so we decided to question an LLM (in this instance, ChatGPT) about why this might be the case. The response we received is as follows: LLM Response 5.1: Your observation about the tone and manner of speech in translations highlights an important aspect of translation quality that COMET might not fully capture. Here’s why: 1. Literal Accuracy vs. Stylistic Fidelity: COMET is primarily designed to evaluate the literal accuracy and fluency of translations. It does well in judging how close the MT is to the reference in terms of meaning and linguistic quality. 2. Stylistic Nuance: While COMET can capture some aspects of fluency and naturalness, it might not fully appreciate stylistic nuances such as the specific tone, voice, or manner of speech that are present in the reference translation. For example, an elderly man’s distinctive speech patterns might not be fully valued in COMET’s scoring if the reference doesn’t explicitly indicate the importance of these characteristics. 3. Contextual and Cultural Sensitivity: Human evaluators can recognise and appreciate subtleties such as tone, cultural references, and stylistic consistency. Automated metrics like COMET are still limited in their ability to assess these elements comprehensively. (It should be noted that such categories are analytical constructs used by human evaluators; the LLM itself does not “understand” or deliberately apply them.) 96 5 Teaching translation with AI This type of activity can provide a novel translator training experience. First, learners understand the relationship between few-shot prompts and parallel corpora, learning that LLMs can generate similar translations from them. While the above example focuses on tone and manner, it could also be interesting to investigate whether LLMs can replicate aspects such as terminology and grammatical structures. As noted in related work, similar attempts have been made in the context of metaphor (Dorst 2024). To achieve this, it might be necessary not only to show parallel corpora but also to add specific prompts that supplement what aspects of the language should be learned from them. Here, the practical application of TS concepts and professional translation considerations can be implemented in the form of prompts. Moreover, this approach encourages critical thinking about translation quality assessment. By comparing human evaluations with automated metrics such as COMET, learners can develop a deeper understanding of the complexities involved in translation evaluation. This specific exercise highlights the importance of considering both quantitative metrics and qualitative aspects such as style, tone, and cultural nuances in assessing translation quality. It also underscores the current limitations of automated evaluation tools and the continued relevance of human expertise in translation assessment. Furthermore, engaging with LLMs to analyse discrepancies between human and automated evaluations can provide valuable insights into the strengths and limitations of different evaluation methods, fostering a more comprehensive and nuanced approach to translation quality assessment in both educational and professional contexts. 6 CoT prompting and the translation process In this section, a simple CoT prompt was used, employing a taxonomy of translation strategies. Rather than recreating the translation process through a CoT prompt, in other words, not a strict step-by-step prompt, it facilitated a relatively high-resolution cognitive process concerning translation. Specifically, the wellknown procedures described by Vinay & Darbelnet (1958/1995) were provided to the LLM. The definitions were generated by ChatGPT-4o. Subsequently, the model was instructed to translate a Japanese sentence (previously used in an example) by applying all the procedures from the provided taxonomy. Let us first examine the following prompt: 97 Masaru Yamada LLM Prompt 5.2: Translate the English sentence “When I was young, life was so different” into Japanese, adhering to the provided taxonomy below. Direct Translation Methods 1. Borrowing: Directly taking words from the source language into the target language without translation (e.g., “pizza”). 2. Calque: Translating a foreign word or phrase literally, creating a new expression in the target language (e.g., “skyscraper” as “gratte-ciel” in French). 3. Literal Translation: Word-for-word translation that adheres closely to the original syntax and meaning, appropriate when the languages share similar structures and cultural contexts. Oblique Translation Methods 4. Transposition: Changing the grammatical category of a word without altering its meaning (e.g., transforming a noun into a verb). 5. Modulation: Changing the form of the message, introducing a shift in point of view or cognitive category to fit the natural expression in the target language (e.g., “It’s not difficult” instead of “It’s easy”). 6. Equivalence: Using an entirely different expression to convey the same situation, often applied in idiomatic expressions or proverbs (e.g., “Out of sight, out of mind” translated as “Loin des yeux, loin du cœur” in French). 7. Adaptation: Modifying the cultural reference when the source language situation is unknown or unrecognisable in the target culture (e.g., changing a reference to Thanksgiving in the U.S. to a relevant local holiday in the target culture). While such taxonomies of translation strategies are often lectured on in introductory TS courses, it was not realistic to use classification tables for discussing translations in practical translation classes. Especially when teaching translation into Japanese, European-developed classifications often did not directly apply, necessitating the use of customised literature or adaptation for practical use. In essence, this strategy classification taxonomy exemplifies how categorisations and accumulations from TS research were seldom utilised in practical teaching classes. Frankly, it was cumbersome. However, this is not to say that human 98 5 Teaching translation with AI translators do not mentally engage in a process similar to this taxonomy, creating and selecting from multiple translation options. Nor is it to suggest that such taxonomies are entirely without merit. To reiterate, the primary issue was the time-consuming nature of the process. Given this context, we provided the aforementioned prompt and requested a translation. The result is as follows. For the translation, Claude 3.5 Sonnet was utilised, consistent with the earlier example. LLM Response 5.2: Borrowing: ヤングの頃、ライフはとても違っていた。 Calque: 若かった時、生命は非常に異なっていた。 Literal Translation: 若かった頃、人生はとても違っていた。 Transposition: 若い時代には、人生は大きく異なっていた。 Modulation: 若かった頃は、今とは全く違う人生だった。 Equivalence: 若い頃は、今とは別世界だった。 Adaptation: 青春時代は、今とは天と地ほど違っていた。 As mentioned earlier, Vinay and Darbelnet’s translation taxonomy is challenging to apply to Japanese translation. Therefore, I was impressed that the LLM could instantly ‘understand’ this taxonomy and create examples adapted to Japanese translations. However, understanding this classification in Japanese remains difficult, so I supplemented the prompt by asking for specific explanations of “which parts of the translation each strategy [this] is applied to”. Consequently, the following detailed explanations were obtained. Due to space constraints, only the first three results are presented here. LLM Response 5.3: 1. Borrowing: “Yangu no koro, raifu wa totemo chigatte ita.”(ヤングの頃、 ライフはとても違っていた。) - The English words “young” and “life” are borrowed directly into Japanese as “yangu” and “raifu” respectively. 2. Calque: “Wakakatta toki, seimei wa hijou ni kotonatte ita.”(若かった 時、生命は非常に異なっていた。) - “Life” is translated as “seimei” (生命), 99 Masaru Yamada which is a more literal translation of “life” as in “living being”, rather than the more common “jinsei” (人生) meaning “lifetime” or “course of life”. - “So different” is translated more literally as “hijou ni kotonatte ita” (非常に異なっていた). 3. Literal Translation: “Wakakatta koro, jinsei wa totemo chigatte ita.”(若 かった頃、人生はとても違っていた。) - The sentence structure of the original is maintained, with each word translated to its Japanese equivalent. … In actual translation practice, these strategies would typically be combined to create a natural and effective target text. Some of these examples may appear unnatural as they are designed to highlight the specific features of each translation strategy. While it may seem repetitive, I must emphasise how remarkably clear, accurate, and well-explained these examples are. The precision and correctness of both the instances and their descriptions are noteworthy. This verification suggests the potential for LLMs to support human translators throughout their translation process using CoT prompting. However, more immediately apparent is how such activities could offer significant insights for translation education. The example presented here, applying a well-known taxonomy of translation strategies to create translation examples, is just one illustration of this potential. Beyond this, we might also consider the application of other TS assets, such as learning about translation issue typologies. These exercises provide a tangible sense of how theoretical concepts from TS can be practically applied. In essence, this approach not only demonstrates the impressive capabilities of LLMs in understanding and applying complex translation theories but also opens up new possibilities for translation education. By bridging the gap between theoretical knowledge and practical application, it offers a novel way to engage with TS concepts that have traditionally been challenging to implement in practical training settings. 100 5 Teaching translation with AI 7 Discussion and concluding remarks Firstly, it is crucial to understand that prompt engineering is not solely aimed at improving automatic translation quality based on NLP technologies or unified value systems. We often perceive AI and LLMs as adversarial to human translators. While important, this perspective, which leads to a focus on cautious educational or practical use – such as copyright concerns or ethical considerations – may be overly conservative. To draw an extreme analogy, it is akin to focusing solely on avoiding accidents or problems when planning an enjoyable trip, rather than considering how to make the journey truly rewarding. This chapter has specifically explored the positive and proactive possibilities of utilising LLMs in translator training. To elucidate these positive applications, we have examined the use of accumulated TS assets and metalanguages as prompts for LLMs. However, I must reiterate that the primary aim of this chapter was not to apply TS concepts to prompts for the simple goal of improving LLM-generated translation quality, as is often the focus in NLP. Instead, the objective was to demonstrate how translation metalanguages can be used to engage in dialogue with AI, potentially enhancing translation education.3 Specifically, we have presented concrete examples of learning scenarios using translation concepts within the frameworks of few-shot prompts and CoT prompting. The prompt utilising process taxonomy to generate various translation variations was particularly surprising and insightful. Critical analysis and AI consultation on this approach yielded highly accurate responses. While these examples are limited and the high accuracy may be partly attributed to the use of the well-known Vinay and Darbelnet taxonomy, they represent just the first step in exploring a vast potential. As an educator, I am both intrigued and obligated to further investigate these possibilities. While these results are promising, it is important to note that this is just the beginning. Further research and experimentation could reveal even more ways in which LLMs can enhance translation education and support professional translation practices. Moreover, this method could potentially make the learning process more interactive and engaging for students, allowing them to explore various translation strategies and their implications in a more hands-on manner. It also provides a platform for discussing the nuances of translation choices, which is crucial for developing critical thinking skills in aspiring translators. This ap3It should be emphasised that such systems do not in fact “understand” translation concepts; rather, they generate output based on statistical predictions of likely word sequences. 101 Masaru Yamada proach may offer opportunities to foster autonomous learning and develop essential qualities for independent translators. It is also important to acknowledge some limitations. Not all responses from LLMs were accurate, and some prompts were less successful than others. For instance, while the LLM correctly explained why COMET could not provide a fair evaluation in the few-shot prompt example, it failed to give a reasonable answer when asked which translation (Claude or DeepL) was closer to the reference translation. Such errors and limitations of LLMs become more apparent with increased use. However, given that translator training inherently requires maintaining a critical perspective, I believe exploring the possibilities of using LLMs is as important as considering the risks. In conclusion, this chapter has demonstrated concrete methods for exploring the potential of LLMs in translation education. By leveraging the concepts and metalanguages of TS in prompt engineering, we can create more engaging, interactive, and effective learning experiences for translation students. While challenges and limitations exist, the potential benefits of integrating LLMs into translator training are significant and warrant further investigation and development. References American Translators Association. 2024. ATA statement on artificial intelligence. https://www.atanet.org/advocacyoutreach/atastatementonartificialintelligence/. Dorst, Aletta G. 2024. Metaphor in literary machine translation: Style, creativity and literariness. In Andrew Rothwell, Andy Way & Roy Youdale (eds.), Computer-assisted literary translation, 173–186. New York, USA: Routledge. DOI: 10.4324/9781003357391-9. ELIS. 2025. ELIS 2025 European language industry survey. Tech. rep. ELIS Research. 1–53. http://elissurvey.org/wpcontent/uploads/2025/03/ELIS2025_Report.pdf. European Council of Literary Translators’ Associations. 2024. No one left behind, no language left behind, no book left behind. https://www.ceatl.eu/no-one-leftbehind-no-language-left-behind-no-book-left-behind. Gao, Yuan, Ruili Wang & Feng Hou. 2023. How to design translation prompts for ChatGPT: an empirical study. https://arxiv.org/abs/2304.02182. He, Sui. 2024. Prompting ChatGPT for translation: a comparative analysis of translation brief and persona prompts. https://arxiv.org/abs/2403.00127. 102 5 Teaching translation with AI Hendy, Amr, Mohamed Abdelrehim, Amr Sharaf, Vikas Raunak, Mohamed Gabr, Hitokazu Matsushita, Young Jin Kim, Mohamed Afify & Hany Hassan Awadalla. 2023. How good are GPT models at machine translation? A comprehensive evaluation.https://arxiv.org/abs/2302.09210. House, Juliane. 1981. A model for translation quality assessment. 2nd edn. Tübingen, Germany: Gunter Narr. ITI. 2024. Slow translation manifesto. https://www.iti.org.uk/discover/policy/ slow-translation-manifesto.html. Japan Association of Translators. 2024. Statement on the public and private sector initiative to use AI for high-volume translation and export of manga. https:// prtimes.jp/main/html/rd/p/000000001.000143535.html. Jiao, Wenxiang, Wenxuan Wang, Jen-tse Huang, Xing Wang, Shuming Shi & Zhaopeng Tu. 2023. Is ChatGPT a good translator? Yes with GPT-4 as the engine. https://arxiv.org/abs/2301.08745. LITHME Project. 2021. Language in the human-machine era. https://lithme.eu/. Miyata, Rei, Masaru Yamada & Kyo Kageura. 2023. Metalanguages for dissecting translation processes: Theoretical development and practical applications. London, UK: Routledge. https : / / www . routledge . com / Metalanguages - for - Dissecting-Translation-Processes-Theoretical-Development-and-PracticalApplications/Miyata-Yamada-Kageura/p/book/9781032168951. Nida, Eugene A. & Charles R. Taber. 1969/2003. The theory and practice of translation. Leiden, Netherlands: Brill. Papineni, Kishore, Salim Roukos, Todd Ward & Wei-Jing Zhu. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, 311–318. Philadelphia, USA: Association for Computational Linguistics. DOI: 10.3115/ 1073083.1073135. https://www.aclweb.org/anthology/P02-1040. Rei, Ricardo, Craig Stewart, Ana C. Farinha & Alon Lavie. 2020. COMET: A neural framework for mt evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP 2020), 2685–2702. Online: Association for Computational Linguistics. DOI: 10.18653/v1/2020.emnlpmain.213. https://www.aclweb.org/anthology/2020.emnlp-main.213. Sayers, Dave, Rui Sousa-Silva, Sviatlana Höhn, Lule Ahmedi, Kais AllkiviMetsoja, Dimitra Anastasiou, Štefan Beňuš, Lynne Bowker, Eliot Bytyçi, Alejandro Catala, Anila Çepani, Rubén Chacón-Beltrán, Sami Dadi, Fisnik Dalipi, Vladimir Despotovic, Agnieszka Doczekalska, Sebastian Drude, Karën Fort, Robert Fuchs, Christian Galinski, Federico Gobbo, Tunga Gungor, Siwen Guo, Klaus Höckner, Petralea Láncos, Tomer Libal, Tommi Jantunen, Dewi Jones, Blanka Klimova, Eminerkan Korkmaz, Sepesy Maučec Mirjam, Miguel Melo, 103 Joss Moorkens & Gökhan Doğru withdrawn with little warning due to a change of circumstances (or geopolitical expediencies). At the time of writing, there is a push for LLM sovereignty, with international organizations like the European Union, national governments and some companies investing in supercomputers and building their own LLMs. Of course, not all LMs have to be huge, but the focus for US big tech providers continues to be scale, creating larger and larger models that take longer to train, with knock-on effects regarding environmental sustainability. The pretraining and deployment of foundational LLMs is a resource-intensive process that requires significant capital investment for acquiring the necessary hardware, training data and human talent as well as operating costs, and each large-scale pre-training cycle has significant environmental impacts due to high electricity and water consumption. Combined, these factors allow only a few companies to have very powerful LLMs. Since many national governments, large corporations and public organizations consider LLMs and AI in general as strategic enablers for their future, they are increasing their investment in this area as well and setting big targets for broader AI use. However, the impact of this AI ‘arms race’ on the environment is likely to be catastrophic on our planet if every entity tries to develop their own proprietary LLM. In recent years, AI systems have in general become more efficient, but the trend for scale means quality improvement is based on more data and parameters, requiring more training time. Luccioni et al. (2025) argue that increased efficiency in AI training and in time saved by using AI will tend to spur more use of AI, continuing an upward trend not only in emissions, but in water use for data centres, rare earth metals used for ICT, and more harmful electronic waste. These environmental harms and the impacts directly from AI are difficult to measure exactly. Emission levels may differ depending on whether energy comes from fossil fuel or renewable sources (Shterionov & Vanmassenhove 2022) or depending on the time of day and data centre location (Dodge et al. 2022). Big tech firms are looking to buy up renewables and nuclear power sources as they come online, in order to maximise their net-zero credentials. Water footprints are also likely to differ depending on time and region (Li et al. 2025). Presently, GenAI does not require a large proportion of resources, but projections, such as those from the head of the UK National Grid predicting an AI-driven six-fold increase in power requirements for data centres in the next decade (BBC 2024), are worrying. For now, energy and water use attributable to GenAI are very small in comparison to the huge requirements for watching streaming media or joining a Zoom meeting (Mytton et al. 2024). 110 6 Teaching AI ethics for translation students 4 Possible solutions Positioned within the international AI ‘arms race’, with politically-motivated support for some developers leading to preferred companies receiving preferential access to governments and cutting-edge technologies, there are also movements to use ‘AI for Good’. One platform by that name is led by the International Telecommunication Union of the United Nations (UN) to use AI to achieve UN sustainable development goals, and there are smaller initiatives such as the Distributed Artificial Intelligence Research Institute that looks to push back against the influence of big tech on AI research, development and deployment. Many people seek to use (Gen) AI to reduce harms and inequality. There have been proposed uses of AI to route power use to maximise the use of renewable energy and to improve energy efficiency in the design, building, and use of commercial buildings (Ding et al. 2024). Van Wynsberghe (2021) feels that there are two motivations behind ‘AI for sustainability’ and ‘sustainability of AI’ that ought to be combined. The former seeks to do good, yet might entail negative environmental impacts, whereas the latter acknowledges that for AI to be sustainable, there needs to be lower environmental costs for AI training, tuning and inference. She defines sustainable AI as a necessary movement to “foster change in the entire lifecycle of AI products (i.e. idea generation, training, re-tuning, implementation, governance) towards greater ecological integrity and social justice” (Van Wynsberghe 2021: 217). National and international legislation, most notably the EU AI Act, seek to limit harmful uses of AI. The AI Act defines a typology of tiered uses of AI based on risk, with ‘unacceptable risk’ uses forbidden and high-risk uses, such as recruitment decision-making and job allocation, subject to special regulation. This renders algorithmic job allocation in the EU illegal, although it’s likely that the recommendation of an automated project management tool will still be followed. In a position paper by Moorkens et al. (2024) we borrowed the idea of a triple bottom line from Business Ethics (Elkington 1997) to propose that translation technologies and LLMs be evaluated not just focusing on performance, but rather giving equal weight to people, planet and performance. This is necessarily a heuristic rather than an exact metric, but follows on from criticisms of a focus purely on performance pushing AI development in the wrong direction by Schwartz et al. (2020) and others. For people, we might consider how the LLM impacts annotators, translators, platform workers, and those who have been dispossessed of their data, balancing these against benefits to people. For the planet, we might look at energy costs/CO2, efficient models, and ICT cost 111 Joss Moorkens & Gökhan Doğru and disposal. Finally, for performance, we should use task-appropriate and comprehensive standards. A previous suggestion for translation data from Moorkens & Lewis (2019) was a community-owned and managed digital commons, following the ideas of Ostrom (2011), with tiered access available for a cost. We argued that this would help to “sustain the occupation of translation and to minimise the potential risks and harms to translators and the public” (Moorkens & Lewis 2019: 17). This idea seems to be similar to the intended implementation of a European Language Data Space as a decentralised marketplace for text, video and audio data in different languages (Rehm et al. 2024). The advice from some researchers, such as Rudin (2019), is to entirely avoid black-box, subsymbolic systems for high-risks uses. This is because opacity is largely baked into subsymbolic AI systems, as described previously, although Rudin and others believe that, in many cases, comparable results may be achieved with more transparent, hybrid systems. For very large, closed source systems, we do not know what training data has been used or what the RLHF guardrails are. However, not all LLMs are closed source – or even that large. We mentioned the alternative options for small-scale LMs in Section 2. These can be useful for narrowly-defined tasks, with benefits of low cost, low environmental impact, and customisability. For technically confident students, guidelines for building a custom small LM are provided by Moorkens et al. (2025). In academic research ethics, the credo has moved on from ‘do no harm’ to the need to actually benefit research participants. Relatedly, best practice for engaged research involves a participatory approach, working cooperatively, particularly with marginalised groups, as co-creators of knowledge rather than imposing narratives or putting words into their mouths. Birhane et al. (2022) propose this approach for building AI systems, offering examples of participatory approaches that aim to lessen existing imbalances of power. In this way, developers “acknowledge that the communities and publics beyond technical designers have knowledge, expertise and interests that are essential to the development of AI that aims to strengthen justice and prosperity” (Birhane et al. 2022: 7). 5 Translation and AI ethics in the classroom 5.1 Classroom discussion activities As student users of translation technology – and most likely users of related AI tools and cloud services more broadly – students in a translation classroom 112 6 Teaching AI ethics for translation students are already part of the interconnected web of ethical issues from the previous section. They may not have given them much thought, as so much of the hype about GenAI focuses on its positive potential rather than ethical issues. GenAI is also entangled with what Brand & Wissen (2021) call the ‘imperial mode of living’, through which public and private organisational strategies and individual lifestyles and practices in the Global North rely on the unlimited appropriation of resources, a disproportionate claim to global and local ecosystems, and cheap labour, ideally from distant locations. Bryan (2022: 330) conceptualises individual relations with the climate crisis as a “form of ‘difficult knowledge’, particularly as it relates to learners’ self-implication in the conditions that are being addressed”. We can reasonably broaden this to AI ethics for trainee translators and educators. Transmission-based lectures alone seem inappropriate for this topic, as difficult knowledge may raise sensitivities; nobody likes to be hectored or to feel that their personal ethics are in question. In this section, we introduce two methods for stimulating discussion and reflection about translation and AI ethics in the classroom. The first uses scenarios or case studies, placing ethical dilemmas into familiar contexts for discussion. The second draws from Bryan’s (2022) ‘pedagogy of the implicated’, which seeks to prompt critical reflection about our own positioning as ‘implicated subjects’ and to foster agency for change. Bryan uses the notion of the implicated subject from Rothberg (2019) to look beyond dichotomies of individual versus institutional responsibility for injustices to a discussion of how we are enmeshed with systems in many ways across historical (diachronic) and contemporary social-structure (synchronic) lines. 5.2 Case studies According to Benbunan-Fich (1998), a combination of lectures and discussion are complementary ways of introducing ethical issues in the classroom. To begin with, lectures about “ethical concepts can lay the theoretical foundations” so that students can “practice ethical analyses” thereafter using case studies (Benbunan-Fich 1998: 20). Case studies have proved to be a useful tool, particularly in business schools, for many years. According to Barnes et al. (1994), case studies extend learning beyond each class, stimulating deeper insights that link across classes and modules. Led by instructors with appropriate case studies, students will engage and can develop and articulate critical insights. These come through four particular factors: situational analysis, active student involvement, a non-traditional instructor role and the need to relate analysis and action. Situational analysis means that ethical issues are applied in situ rather than in 113 Joss Moorkens & Gökhan Doğru theory. Students are active, engaged and learning without direct teaching, energised and challenged by discussions. This puts the instructor in a non-traditional role of facilitator without dominating the classroom. Rather than only building knowledge, case studies combine analysis and action: the “importance of action influences the entire case discussion, which focuses on the practical and doable” (Barnes et al. 1994: 48). The case study chosen must be appropriate – Zhou (2022: 400), for example, recommends finding “representative translation-ethical cases”. There are two that may be used or adapted to GenAI from Moorkens (2022) in the open access book Machine Translation for Everyone. Case studies may be accompanied with a general question to motivate discussion (e.g. ‘What ethical issues do you see in this scenario?’) or a case worksheet featuring a number of questions or tasks. Some examples that could be included are: • Describe the key ethical issues and the stakeholders involved. • What ethical theories might provide guidance in this scenario? • Can you think of guidelines from a translator code of ethics that could be useful in this scenario? • Are there any legal issues involved in this scenario? • With the above questions in mind, are there any actions that you propose to resolve this scenario in an ethical manner? Students may address these questions in groups or as part of a plenary discussion moderated by the instructor. A final recap, making note of any new or unexpected issues that were brought up by students, can effectively summarise a class and reinforce key lessons. 5.3 Pedagogy of the implicated The approach proposed by Bryan (2022) in her article on ‘pedagogy of the implicated’ could work in combination with or instead of case studies. As mentioned, Bryan draws on Rothberg (2019) for her description of (most of) us as “implicated subjects”, who can be “beneficiaries and perpetuators of systems that are not of [our] own making or that [we] have no direct ancestral attachment to” (Bryan 2022: 338, italics in original). A preamble explaining this complex entanglement of implication is a valuable precursor to discussion in the classroom. The prompt 114 6 Teaching AI ethics for translation students for this discussion could come from carefully chosen images or a mapping exercise to visualise these entangled relationships. Bryan’s (2022: 341) article is focused on climate change education, for which there is a tendency towards what she calls an “apocalyptic sublime aesthetic”. This aesthetic, she argues, “positions viewers as mere voyeurs or passive spectators – rather than active agents or implicated subjects – in the unfolding chaos”. Images of AI are perhaps generally less apocalyptic, but are similarly uninvolving. These science-fiction-inspired images will be familiar to those who have seen many visualisations of AI featuring stylised robots with fingers to their chins among gigantic circuit boards, concealing any human involvement, ethical dilemmas, or even any question about the smooth transition towards wise and sentient intelligent robots. Dihal & Duarte (2023) summarise how typical stock images of AI are misleading for non-experts, in that they hide societal and environmental impacts, promote unrealistic expectations of AI capabilities, hide human accountability, and often promote stereotypical assumptions about gender, ethnicity and religion. This led them to start the website Better Images of AI1featuring artist-created images, collages and illustrations that foreground the issues hidden in regular stock images. These images are highly considered and thus make an excellent starting point for classroom discussion and reflection. Bryan proposes a mapping exercise as part of a discussion of self-implication in a complex interconnected consumer-focused system, such as that behind GenAI, to prompt critical self-reflection. The main example that she uses is what she calls the Social Ecology of Responsibility Framework (SERF), inspired by Bronfenbrenner’s (2009) theory of the ecology of human development. Bronfenbrenner was “concerned with the mechanisms, processes and conditions that shape individuals’ development and devised a model that theorises the reciprocal interactions that occur within and between different nested environments (or systems) which in turn affect developmental outcomes” (Bryan 2022: 338). The SERF can be a useful tool to visualise the interconnections between the individual actions of the ‘implicated subject’ within their immediate context and larger social, national and international systems. These are represented by concentric circles, expanding from the individual’s microsystem (our immediate context) to our macrosystem of societal, political, and cultural norms and context. Within these microand macrosystems are the mesosystem, describing points of connection between different levels, and the exosystem of institutions and organisations that exert influence without any direct contact with us as individuals. 1https://betterimagesofai.org 115 Joss Moorkens & Gökhan Doğru Navarro & Tudge (2023) suggest that our microsystems should be subdivided into physical and virtual categories, as our day-to-day activities, social roles, and interpersonal relations may differ (or blur) between physical and digitallymediated domains. The mesosystem groups our multiple microsystems together. We might consider the remarkably consistent messaging about technology and GenAI coming from the exoand macrosystems, why that might be, and the potential impacts on our mesosystem. Figure 1 features an adaptation of Bryan’s (2022) SERF, using a layered funnel design and bringing in the idea of the triple bottom line. This could form a useful starting point for classroom discussions or an example for a mapping exercise. The intention is not to blame individuals – the chronosystem part of the SERF makes clear the historical context for current systems – but to empower the individuals at the centre, which of course include the educators as elements and perpetuators within the overall system. Figure 1: An adapted version of Bryan’s (2022) Social Ecology of Responsibility Framework with some ideas to use for classroom discussion Another useful image intended to stimulate reflection and discussion of the 116 6 Teaching AI ethics for translation students translator’s responsibilities, with the potentially conflicting pressures from ethical codes, business ethics, personal interests, and social responsibility comes from Joseph Lambert’s (2023) book Translation Ethics. The section on responsibilities and related questions place many of the points from this chapter within the context of translation work. 6 Conclusion Many curricula now incorporate AI literacy as an important competence. Chan & Colloton (2024) recommend a fine-grained interpretation of AI literacy that is tailored to subject or professional areas, and a major part of this for translation students is AI ethics in translation. In this chapter, we briefly introduce some of the issues and debates regarding AI ethics, along with some positive initiatives that are ongoing. We propose interactive, social constructivist methods for teaching students about AI ethics for translation, highlighting two potential methods in particular. The first of these is the use of carefully chosen case studies with accompanying support material to help students talk through ethical issues. Benbunan-Fich (1998: 21) feels that it is “crucial for students to practice ethical decision making, so when they have ethical decisions to make in the real business world they have a framework to follow”. The second suggested method is an adaptation of Bryan’s (2022) pedagogy of the implicated, using images and mindmapping to visualise and understand our own place within the tangle of beneficiaries and perpetrators of the system within which GenAI is developed and marketed. A disingenuous tendency to highlight individual responsibility for issues of ethics and sustainability lets large organisations off the hook, but these methods show how individual and collective responsibilities are interlinked. Our status as implicated subjects means that we must also take responsibility for our actions and ethical decisions. We hope that these methods will be effective beyond the individual class, leading to deeper knowledge in students that can come to their own individual ethical position on the debates and issues in AI ethics for translation. Acknowledgement This was written with the financial support of Research Ireland at ADAPT, the Research IrelandCentre for AI-Driven Digital Content Technology at Dublin City University [13/RC/2106_P2]. 117 Joss Moorkens & Gökhan Doğru References Abdallah, Kristiina. 2011. Towards empowerment: Students’ ethical reflections on translating in production networks. The Interpreter and Translator Trainer 5(1). 129–154. DOI: 10.1080/13556509.2011.10798815. Barnes, Louis B., C. Roland Christensen & Abby J. Hansen. 1994. Teaching and the case method: Text, cases, and readings. 3rd edn. Boston, USA: Harvard Business School Press. BBC. 2024. Data centre power use ‘to surge six-fold in 10 years’. https://www.bbc. com/news/technology-68664182. Benbunan-Fich, Raquel. 1998. Guidelines for using case scenarios to teach computer ethics. ACM SIGCAS Computers and Society 28(3). 20–24. DOI: 10.1145/ 298972.298982. https://dl.acm.org/doi/10.1145/298972.298982. Bender, Emily M. 2011. On achieving and evaluating language independence in NLP. Linguistic Issues in Language Technology 6. DOI: 10.33011/lilt.v6i.1239. Birhane, Abeba, William Isaac, Vinodkumar Prabhakaran, Mark Diaz, Madeleine Clare Elish, Iason Gabriel & Shakir Mohamed. 2022. Power to the people? Opportunities and challenges for participatory AI. In Proceedings of the 2nd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ’22). Arlington, USA: Association for Computing Machinery. DOI: 10.1145/3551624.3555290. Boyle, Siân. 2025. Revealed: Google facilitated Russia and China’s censorship requests. The Observer. https://www.theguardian.com/world/2025/feb/15/ google-helped-facilitate-russia-china-censorship-requests. Brand, Ulrich & Markus Wissen. 2021. The imperial mode of living: Everyday life and the ecological crisis of capitalism. New York, USA: Verso. Bronfenbrenner, Urie. 2009. Ecology of human development: Experiments by nature and design. Cambridge, USA: Harvard University Press. Brown, Peter, John Cocke, Stephen Della Pietra, Vincent J. Della Pietra, Frederick Jelinek, Robert L. Mercer & P. Roossin. 1988. A statistical approach to French/ English translation. In Proceedings of the Second Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages. Pittsburgh, USA. https://aclanthology.org/1988.tmi-1.19/. Bryan, Audrey. 2022. Pedagogy of the implicated: Advancing a social ecology of responsibility framework to promote deeper understanding of the climate crisis. Pedagogy, Culture & Society 30(3). 329–348. DOI: 10.1080/14681366.2021. 1977979. https://www . tandfonline.com/doi / full/10.1080 / 14681366.2021. 1977979. 118 6 Teaching AI ethics for translation students Chan, Cecilia Ka Yuk & Tom Colloton. 2024. Generative AI in higher education: The ChatGPT effect. London, UK: Routledge. Dihal, Kanta & Tania Duarte. 2023. Better images of AI: a guide for users and creators. Tech. rep. Cambridge, UK: The Leverhulme Centre for the Future of Intelligence, We & AI. https://blog.betterimagesofai.org/wp-content/uploads/ 2023/02/Better-Images-of-AI-Guide-Feb-23.pdf. Ding, Chao, Jing Ke, Mark Levine & Nan Zhou. 2024. Potential of artificial intelligence in reducing energy and carbon emissions of commercial buildings at scale. Nature Communications 15(1). 5916. DOI: 10.1038/s41467-024-50088-4. https://www.nature.com/articles/s41467-024-50088-4. Dodge, Jesse, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A. Smith, Nicole DeCario & Will Buchanan. 2022. Measuring the carbon intensity of AI in cloud instances. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22), 1877–1894. Seoul, Republic of Korea: Association for Computing Machinery. DOI: 10.1145/3531146.3533234. Elkington, John. 1997. Cannibals with forks: The triple bottom line of 21st century business. Stony Creek, USA: New Society Publishers. Ferrero, Tania, Leonor Pais & Nuno Rebelo dos Santos. 2015. Decent work: An aim for all made by all. International Journal of Social Sciences 4(3). 30–42. https: //www.eurrec.org/ijoss-article-193. Fırat, Gökhan. 2024. Translation work in the digital economy: Working conditions of translators on platforms and in cooperatives. University of Surrey. (Doctoral dissertation). https://openresearch.surrey.ac.uk/esploro/profile/g%C3% B6khan_firat/output/all?institution=44SUR_INST. Giroux, Henry A. & Peter McLaren. 1997. Paulo Freire, postmodernism and the utopian imagination: A Blochian reading. In Jamie Owen Daniel & Tom Moylan (eds.), Not yet: Reconsidering Ernst Bloch, 138–162. London, UK: Verso. Kelleher, John D., Brian MacNamee & Aoife D’Arcy. 2015. Fundamentals of machine learning for predictive data analytics: Algorithms, worked examples, and case studies. Cambridge, USA: MIT Press. Lambert, Joseph. 2023. Translation ethics. London, UK: Routledge. DOI: 10.4324/ 9781003148265. Li, Pengfei, Jianyi Yang, Mohammad A. Islam & Shaolei Ren. 2025. Making AI less “thirsty”: uncovering and addressing the secret water footprint of AI models. arXiv:2304.03271 [cs]. DOI: 10.48550/arXiv.2304.03271. http://arxiv.org/abs/ 2304.03271. 119 Senem Öner Bulut 1 Introduction The impact of machine translation (MT) on the translation profession and translator education has been profound in the last decade, as has the research into this impact. The literature has focused prominently on the tense relationship between the human translator and the machine/technology, which Pickering (2008) conceptualised as a “dance of agency”. Olohan (2011: 354) drew attention to the “decentring” of the human agent in this context. O’Brien (2012: 119–120) further characterised translation as “human-computer interaction”, defining what was happening then as a “shift in paradigm” and pointing out the potential for “dialogue” between human translators and technology. Ruokonen & Koskinen (2017: 321) further elaborated on the human-machine dynamic, with a focus on the emotional dimension inherent in this relationship. They found that the human agents were both “willing” and “reluctant” in terms of assigning agency to the machine and that the relationship between the human and the machine is too “complex” to be reduced to “simplified man versus machine conceptions”. In a similar vein, Cadwell et al. (2018: 312) observed that human translators held both positive and negative perceptions of MT. Positive perceptions related to reduced effort and increased inspiration while the negative were linked to concerns about the negative influence of MT on human creativity. The authors also argued that the ‘dance of agency’ between the human and the machine was a symbiotic one, as they mutually feed off and shape each other; the “distinction between the human and the material agent” is blurred (Cadwell et al. 2018: 303). The practical consequences of the pervasive integration of MT technologies into the global translation industry and their economic impact have also been explored. Moorkens (2017), for instance, noted the negative consequences of the freelance employment model, which has been increasingly adopted in the global translation industry, in terms of prices and isolation. According to him, the focus should be on how the human translator actually works with MT. Similarly, Vieira (2018: 15) found that, rather than feeling an “intrinsic negativity to the technology”, the perceived threat from technology among translators appears to stem from business practices in the industry. Another focus in the related literature has been on the need to redefine the profile of human translators. Sakamoto (2019: 68–69) underlined the necessity of the “redefinition of the concept of ‘translator’, and consequently ‘translation’ per se” on the grounds that “the notion of ‘translation’ is now being challenged by the growth of technologies”. The author also noted that attention should be paid 126 7 Re-positioning the human translator to the uncertainties related to the legal requirements in MT use, pricing for postediting, and post-editors’ profiles and skills (Sakamoto 2019: 63–66). Based on a comprehensive analysis of the emerging needs and requirements in the translation sector, which was conducted with the aim of addressing the “distance between teaching and professional practice” in the context of MT competences, Gaspari et al. (2015: 333–334) reported “a strong need for an improvement in quality assessment methods, tools and training”, underlining the “growing uptake of MT and the perceived increase of its prevalence in future workflows”. There has been extensive and substantial research into embedding in translator education the new skills, abilities and competencies that human translators should possess in the age of MT, as well as into the concomitant need to re-profile translator education. For instance, Doherty & Kenny (2014: 296–297) designed an SMT syllabus, which foregrounds the “empowerment” of the human translator by enabling him/her to intervene in and add value to the SMT workflows when facing legal, ethical and technical obstacles (Kenny & Doherty 2014: 288). This “empowering” approach has been further extended by Mellinger (2017: 284), who proposes a holistic, “cross-module or cross-curricular integration” of post-editing and MT “to avoid compartmentalisation of various competences and skillsets”. Mellinger also suggests the inclusion of “controlled authoring, terminology management, engine tuning, and post-editing” in translation practice courses, as these are “representative of many of the skills mentioned in recent works on machine translation pedagogy” (Mellinger 2017: 284). Pym (2019) drew attention to the identification of “automation-resistant skills” and their integration into translator education. The measures to be taken by translator educators to empower students and help them become “aware of their usefulness in order to maximise their agency as translators” has also been the focus in a study by Moorkens (2018: 375–376), who designed a translation evaluation exercise to enable students to “demystify NMT output”. Based on a reconceptualisation of translator competence in the age of MT, Öner Bulut (2019: 3) designed a learning practice to help students “raise their awareness of their professional self-concept as human translators” and suggested the consideration of human translator competence and human translation meta competence. Nitzke et al. (2019: 248) proposed a novel “post-editing competence model”, which included the core competences of risk assessment competence, strategic competence, consulting competence and service competence. Öner & Öner Bulut (2021: 100) investigated translation students’ perceived difficulties and benefits in the context of “post-editing oriented neural machine translation error annotation and quality evaluation”. 127 Senem Öner Bulut Exploring the “dynamics of the human-machine dance in the translation classroom”, Öner Bulut & Alimen (2023: 377) designed a learning experiment to enable the students to carry out emergent MT-related tasks of post-editing, pre-editing and error annotation. The authors reported that the experiment helped the majority of participating students “raise awareness of their self-concept as human agents and of the human added value they can create while dancing with the machine” (Öner Bulut & Alimen 2023: 387). These examples of past studies focusing on the impact of MT on the translation profession and translator education show that translation studies (TS) research has indeed come a long way to come to terms with and adapt to the paradigm shift (O’Brien 2012) induced by the increasingly all-pervasive technologisation of translation, especially since the launch of NMT in 2016. Yet, the year 2022 presented both translators and translator educators with another, unprecedented challenge. With the launch of ChatGPT in 2022 (OpenAI), which allows for the production of translation through a prompt-based dialogue between human and generative artifical intelligence (GenAI), the early characterisations of translation as human-computer interaction (O’Brien 2012) and dance of agency (Olohan 2011) have become realities in the fullest sense. This poses new questions and opens new areas of research for the TS community, especially for educators and researchers of translator education. In their above-cited research, Öner Bulut & Alimen (2023: 389) noted that, in the face of the rapid advancements in MT technologies and the uncertainties concerning their integration into education, the re-positioning of translator educators is as necessary as that of human translators. It has also been argued that such re-positioning involves an “awareness of the fact that the translator educator has the primary task of diagnosing and promoting emergent areas of human added value creation in the MT age”, and that this can only be achieved by restructuring the translation classroom as a “platform of collaborative learning, where all involved can work together to discover what remains and will remain human in the MT age” (Öner Bulut & Alimen 2023: 389). In the face of the paradigm shift induced by the introduction of ChatGPT and other GenAI tools, translator educators should once again work in collaboration with students to explore the dynamics of integrating these tools into their workflow. Motivated by this very need, the present chapter presents the results of a collaborative research experiment designed to explore the dynamics of the human-GenAI dialogue and find the pathways to be followed in incorporating the insights gained into translator education. The remainder of this chapter is structured as follows. First, the unprecedented practical and theoretical challenges posed by GenAI technology in terms of the 128 7 Re-positioning the human translator very raison d’être of the human translation are explored. Second, information on the design and implementation of the experiment is given. Third, the data obtained from the experiment is analysed. Fourth, a discussion of the results is presented together with conclusions and suggestions for future research. 2 Human added value and the human translator’s self-concept in the age of GenAI: Re-visiting the interpretive task and responsibility of the human translator ChatGPT is a GenAI tool powered by a large language model (LLM) and has not been developed specifically for interlingual translation. However, the revolution it has brought to translation is of great significance as the tool takes prompts from human agents and hence allows them to produce a translation by providing the tool with contextual information and guidance, i.e. by performing prompt engineering (Yamada 2023 and Chapter 5 of this book). This was certainly not the case with MT, which makes GenAI technology a real breakthrough that now requires translator educators to ask and try to answer new questions about human-technology interaction. In my opinion, in this endeavour we should not attempt to reinvent the wheel. Instead, we should revisit the relevant existing insights into MT, especially those that deal with the integration of MT into translator education, in a way that empowers human translators to create human added value, as outlined in the previous section (see Introduction). This should also involve trying to answer the challenging question posed by GenAI: what is human in translation? This is because, unlike MT, GenAI creates the illusion of having an actual dialogue with a non-human agent which pretends to understand, interpret and produce a linguistic message. This new form of challenge should urge us to revisit the interpretive task and responsibility of the human translator. Accordingly, the approach presented here is performing a backward reading. Rather than giving priority to the investigation of the performance, potential and/or limitations of GenAI, what the present study proposes is the instrumentalisation of the human-GenAI dialogue in order to try to reconsider what is/will remain human in translation in a new, brighter light and find ways to incorporate it into translator education in the age of GenAI. In a seminal contribution by Massey & Kiraly (2019) titled “The Future of Translator Education: A Dialogue”, Kiraly states that “transcoding (the mechanical replacement of linguistic units from a list with corresponding units from a parallel 129 Senem Öner Bulut list) is not at the heart of translation at all” (16, emphasis mine) and that, although in the future the translation profession will be “different”, it is “certainly not going to disappear – unless at some point in our evolution we no longer need to interpret texts” (16, emphasis mine). It is no coincidence that Kiraly highlights the act of interpreting as the raison d’être of the human translator, which echoes the hermeneutical approach applied to translation by Schleiermacher (1813/1977). According to Hermans (2015: 101– 102), Schleiermacher believed that the sole option for a translator is “to act as the hermeneuticist does: to work to attain the best possible understanding of the foreign text which nevertheless remains foreign, and present to the reader […] exactly that understanding”. Accordingly, the meaning of a text is not encoded in a text to be decoded by the translator, rather it should be interpreted by the translator to attain the best possible understanding, which constitutes the hermeneutical task of the translator. The translator is then required to present this understanding to the reader, which, according to Herman’s reading of Schleiermacher, “exceeds hermeneutics” (Hermans 2015: 99). In a similar vein, Şerban & Larisa (2016: 295) underline that in Schleiermacher’s approach, the meeting between the author and the reader is “mediated and orchestrated by the translator” and the translator assumes “the most active mediating role”. Although the hermeneutical approach is of critical importance in that it allows the translator room for active mediation, a specific conceptualisation by Arrojo (1997: 18) of the “inevitable interference” of the translator in the context of translation ethics could further the attempt to reconsider what is human in translation. Arrojo views translation ethics in terms of the relationship between the “authorial power” of the translator and his/her “ethical responsibility” (Arrojo 1997: 18). Based on a postmodern view of “language and subject”, the author underlines that “[as] no reading can ever aspire to repeat or protect someone else’s text, translators necessarily have the right to exercise their authorial power ‘as long as their game is played up front’” (Berman 1995: 93, cited in Simon 1996: 36). In the view of the author, “[s]uch ‘right’ implies, however, an ethical responsibility which parallels that of ‘original’ authors” (Arrojo 1997: 18). In the wake of the emerging paradigm brought about by GenAI, this line of thought can be followed by researchers and translator educators to foreground the human translator’s interpretive task and the accompanying responsibility with the aim of exploring human added value (Massey 2021) and human translators’ self-concept (Kiraly 1990, 2000b, Haro-Soler & Kiraly 2019 and Massey & Kiraly 2019) in the age of GenAI. Drawing careful attention to the human aspect of added value, Massey (2021: 39) characterises human added value as being “manifest in translational decision130 7 Re-positioning the human translator making and problem-solving on a conceptual level that transcends the surface lexical realisations by which meaning is conveyed in source and target texts”. He also observes that “the added value that human translators can and do bring to bear resides in their socio-cultural, socio-technical 4EA (Embodied, Embedded, Enactive, Extended and Affective) cognition, as opposed to disembodied, decontextualised artificial intelligence” (Massey 2021: 39). In doing so, the author revisits Venuti’s (2019) advocacy of the “hermeneutical model of understanding translation not as ‘the reproduction or transfer of an invariant that is contained in or caused by the source text’ but as ‘an interpretive act that inevitably varies source-text form, meaning, and effect’” and Pym’s (2003) “minimalist definition of translation competence as ‘the ability to generate a series of more than one viable target text (TTI, TT2 ... TTn) for a pertinent source text (ST); the ability to select only one viable TT from this series, quickly and with justified confidence’” (Massey 2021: 51). The delineation by Massey (2021: 39) of the added value of the human translator vis-à-vis AI strongly captures the conceptualisation of the human translator as an agent who has the hermeneutical task and power of interpreting the text and who bears the responsibility of their interpretive act. This is something that remains quite impossible for the “disembodied, de-contextualised artificial intelligence”. In the age of GenAI, we must therefore pay unprecedented attention to the translator’s “self-concept” (Haro-Soler & Kiraly 2019), which includes “the image of the translator’s social role”, “the translator’s appraisal of his or her competency for translating a particular text” and “an understanding of responsibility towards the other actors in the translation context of situation” (261, emphasis mine). Consequently, we believe that the constructs of human added value and of (human) translator’s self-concept should constitute the privileged foci in exploring the human-GenAI dialogue and incorporating the gained insights into translator education in a way that helps students raise their self-concepts (HaroSoler & Kiraly 2019: 261–262) as human translators whose raison d’être is to perform the hermeneutical task of interpreting and bearing responsibility. It is with this focus in mind that the following collaborative research experiment was conducted. 3 Design and implementation of the collaborative research experiment The research experiment, the results of which are reported in the present study, was designed following the tenets of the collaborative research model, imple131 Senem Öner Bulut mented by Haro-Soler & Kiraly (2019). It follows the social constructivist and emergentist approach to translator education (Kiraly 2000a: 256), whereby the educator-researcher works collaboratively with translation students “to investigate topics in the domain of Translation Psychology with the goal of having teacher-researchers learn about the translator’s psychological ‘self’ right along with their students”. Accordingly, the research experiment presented here was collaboratively performed by graduate level translation students, who participated in the research on a voluntary basis, and a translator educator (also the author of the present chapter). The aim was to explore the dynamics of humanGenAI interaction as human translators work with GenAI tools as prompt engineers. The student-researcher group consisted of five students enrolled in the PhD programme in translation studies at Yıldız Technical University, Türkiye. All had previous translation education and experience and near-native English proficiency. The educator-researcher was an associate professor of translation studies at the same university. In the research process, the student-researchers and the educator-researcher collectively decided on the design of the task of prompt engineering in three sessions held online in June 2024. After the design was agreed upon, the studentresearchers performed the task individually and reported on their learning/ research processes through self-reflexive accounts of their experiences. The initial decisions made in the recorded online discussions concerned the selection of the source texts and of the GenAI tool to be used to translate the texts through crafting and, when necessary, curating prompts (i.e. prompt engineering; Yamada 2023) so as to pre-process, initiate, produce, edit and evaluate translations that achieved the intended quality level and functions. It was collectively decided to use the free version of OpenAI’s ChatGPT. As for the source texts, two English texts, an essay and a book description, were selected on the grounds that both texts were of mixed forms between expressive and operative text types, while also displaying the features of an informative text type (Reiss 1981: 124–125). Further, the translation of both texts demanded meticulous attention to cultural nuances and context, idiomatic expressions, tone, intent, and linguistic style. The first source text was an essay by Anna Quindlen, published in her “Life in the 30s” column in The New York Times (Quindlen 1987). The second source text was the book description of The Time Regulation Institute, the English translation of Ahmet Hamdi Tanpınar’s Turkish novel titled Saatleri Ayarlama Enstitüsü, translated by Alexander Dawe and Maureen Freely and published in 2014 (Tanpinar 2014). Next, decisions had to be made concerning the guidelines, namely the two translation briefs and the general principles of prompt engineering. First, the 132 7 Re-positioning the human translator educator-researcher provided the student-researchers with the existing (albeit limited, due to the fact that it is a relatively new topic) literature on prompt engineering. This led to an exchange of opinions in the first two online sessions. Drawing on those discussions, the educator-researcher then drafted the guidelines for the task and sent them to the student-researchers for evaluation. In the third session, the draft guidelines were discussed and finalised collectively. The educator-researcher and the student-researchers then had to decide whether to use the finalised translation briefs as the first prompt in the dialogue with the GenAI tool. Taking into consideration the results of the experiments already conducted in other studies on prompt engineering (e.g. He 2024, Peng et al. 2023, Yamada 2023), it was decided not to include the translation briefs in prompt engineering and instead, to use them as guidance for human translatorsprompt engineers only. In this way, the human translators were entrusted with the responsibility of the translation act despite performing this act through prompt engineering. The translation briefs for the first and second source texts were as follows: (1) Translation Brief 1: “The Name is Mine” is an essay by Anna Quindlen published in her “Life in the 30s” column in The New York Times on March 4, 1987. An online magazine website editor in Turkey needs a Turkish translation of the essay and asks you to translate it into Turkish to publish it on the website. The intended audience for the translation is primarily adult readers who would be interested in the works of Anna Quindlen, who writes primarily on feminism and family life and is a Pulitzer Prize-winning columnist, journalist and author. The editor asks you to translate the second and ninth paragraphs of the essay as a test. (2) Translation Brief 2: The source text is the book description of the novel The Time Regulation Institute, which is the English translation of the prominent Turkish novelist and poet Ahmet Hamdi Tanpınar’s Turkish novel titled Saatleri Ayarlama Enstitüsü (1961), translated by Alexander Dawe and Maureen Freely and published by Penguin Classics in 2014. A Turkish literary scholar is asking you to translate the English book description into Turkish in order to examine how the English translation of the book (The Time Regulation Institute) is marketed to the English-speaking audience by Penguin Classics. 133 Senem Öner Bulut As for the general principles of prompting, “persona prompt” (He 2024) was selected as prompting method. This meant that prompting began with the statement “You are a professional translator” without specifying any specific domain. It was also decided that prompt engineering would include the following steps: 1) prompting the tool to analyse the source text for the purpose of translation, 2) prompting the tool to translate the source text, 3) prompting the tool to edit its translation through curating prompts where deemed necessary, and 4) prompting the tool to evaluate the quality of, and comment on, its translation. In order to foreground the responsibility of the human translator, a final step, not included within the scope of the prompt engineering task, was added: post-editing the obtained translation output through prompt engineering to produce the translation end-product. Other than these, no specific prompt statements were drafted in order to allow the human translators-prompt engineers some freedom and creativity in their individual dialogue with the GenAI tool. As a result, each source text generated five versions of prompt engineering and five translation end-products. After completing the task, each student-researcher wrote a report on their prompt engineering experience, the analysis of which is presented below. 4 Analysis of the results: Towards deeper awareness of the power of the human translator vis-à-vis GenAI After the experiment was completed, the student-researchers wrote retrospective process reports on their experiences working with GenAI by answering the questionnaire drafted by the educator-researcher. Globally, their responses indicated that the collaborative research had allowed them to both evaluate the potential impact of GenAI on the translation profession and re-affirm their role as human agents and experts in translation. The answers to Question 1 (“What is your general assessment of the experiment?”), indicated that most of the participants came to recognise that human intervention is still needed in several respects. One participant (P4), for instance, noted that during the experiment she “came to the realisation that human translators still have the final say”. Likewise, another participant (P3) stated that despite the “general conception” that LLMs “are already adequate for daily use or even technical translation at some point”, “they would never be able to replace human translators” in terms of “literary texts” and “mixed forms”. Another participant (P1) stated that human translators’ post-editing is “a necessity rather than a preference”, especially for “restructuring expressive language” and “transferring 134 7 Re-positioning the human translator the style”. Participant 2 highlighted her frustration at the AI tool needing “continuous prompting” to edit its translation “until a final translation, which still need[ed] post-editing” was obtained. Another participant (P5) characterised the experiment as being “educating”, noting that although she had used “ChatGPT on various occasions”, the experiment was the first time that she “actually sat down and tried to generate a translation through prompting”. The answers to Question 2 (“Upon completing the experiment, have you faced any emotional challenges as a human agent collaborating with GenAI to produce a translation? If yes, what were these challenges?”) showed that most of the participants faced emotional challenges while working with GenAI. A feeling of frustration was raised by two participants. Participant 1 stated that he got “frustrated” when the LLM did not “directly implement” his suggestions and he needed to “provide prompts again”. Another participant (P5) also stated that she “found talking to ChatGPT quite frustrating” and that she realised that she didn’t “have the required patience to guide it through everything”. One participant (P2) noted that it was “emotionally exhausting to prompt AI for non-literal translation”, especially “if a word is used figuratively” as it was “challenging to explain the function of the word and design the prompt so as to reach an acceptable translation”. Only one participant (P3) described his feeling as “constant puzzlement” rather than “frustration or anger” as the tool required “an in-depth thought process on how to write the prompts” and that he “felt like the LLM wouldn’t be able to complete the tasks” the way he “wanted it to do”. The emotional challenge expressed by one specific participant (P4) differed from the challenges experienced by the other participants in that the “feeling of stress” experienced by this participant stemmed from her underestimation of the “depth of AI’s analytical capabilities” and its “speed”. The participant (P4) also stated she “experienced a lack of confidence while providing prompts” in the translation of the first source text and “became more comfortable” in the translation of the second source text as she familiarised herself with the tool. The answers to Question 3 (“Upon completing the experiment, what have you discovered about your power vis-à-vis AI and your self-concept as a human translator?”) revealed that the participants addressed various aspects of the power of the human translator on the basis of their experience working with GenAI. One participant (P3) described his experience by saying that it felt like being “[a] supervisor trying to teach a very intelligent new employee how to conduct the job and how it can improve itself in certain areas”. According to him (P3), the tool can “easily analyse the text” and “find out what kind of strategies could be needed to translate the text for the [target] audience”. However, it fails to “produce translation strategies to overcome translation problems” as it remains “too ‘faithful’ 135 Atsushi Mizumoto the latter referring to an approach that views learners’ use of multiple languages as a resource rather than a problem. The deficit-oriented perspective, often aligned with traditional language teaching methodologies, perceives L1 or MT use as detrimental to L2 acquisition. This perspective typically leads to policies discouraging or prohibiting MT use, viewing it as an impediment to authentic language learning. With this perspective, it is believed that that MT use may hinder the development of critical thinking skills in the target language and interfere with immersive language experiences. Conversely, the translanguaging view adopts a more inclusive approach to language learning. This perspective values learners’ entire linguistic repertoire as a resource (Wei 2018). Advocates of translanguaging argue that MT can be a valuable tool for accessing and leveraging learners’ full range of linguistic knowledge, potentially enhancing both language awareness and learning outcomes. This view aligns with contemporary understandings of bilingualism and multilingualism, which conceptualise languages as part of an integrated communication system rather than as separate entities. These contrasting perspectives on MT use in language education are further exemplified in the approaches educators and institutions take when addressing student use of MT. Jolley and Maimone’s (2022) comprehensive review of three decades of MT research in language teaching and learning highlights two distinct approaches: the “MT as Cheating” approach, which leads to a Detect-ReactPrevent Response, and the “MT as Resource” approach, which encourages an Integrate-Educate-Model strategy. The “MT as Cheating” perspective, aligned with the deficit-oriented view, treats MT use as a form of academic dishonesty. This approach focuses on strategies to detect unauthorised MT use, react punitively, and prevent future occurrences. Proponents of this view recommend implementing clear syllabus policies against MT use, designing assignments resistant to MT use, and educating students about the pitfalls of relying on MT. This perspective often leads to policies that ban MT use outright, viewing it as incompatible with language learning goals. In contrast, the “MT as Resource” approach, more closely aligned with the translanguaging view, sees MT as a potential tool for language learning. This perspective advocates integrating MT into the curriculum, educating students on its appropriate use, and modeling effective strategies for leveraging MT in language learning. Researchers like Stapleton & Kin (2019) and Niño (2020) argue for accepting the reality of MT use and finding ways to incorporate it meaningfully into language education. This approach acknowledges the ubiquity of MT in modern life and seeks to prepare students to use it critically and effectively. 238 12 Embracing machine translation in L2 education: The shift from the Detect-React-Prevent mindset to the Integrate-EducateModel approach reflects a growing recognition of the inevitability of MT use in language learning contexts. As Ducar & Schocket (2018) note, the key question is no longer whether teachers can prevent learners from using MT, but rather how to help them use it ethically and effectively as part of their language learning journey. These contrasting approaches to MT use in language education exemplify the broader ideological tensions identified in Grieve et al. (2024) and reflect the outermost layer of ideological factors in Jiang et al.’s (2024) framework. They demonstrate how deeply held beliefs about language acquisition and the role of technology can shape educational policies, pedagogical practices, and ultimately, students’ engagement with and perceptions of MT in their L2 development process. This interplay between ideological stances and practical approaches underscores the complexity of integrating MT into language education and highlights the need for context-sensitive strategies that consider both the potential benefits and challenges of MT use in L2 learning and teaching. 2 Integrating MT into L2 education: A new paradigm 2.1 MT as augmented L2 competence The translanguaging perspective and “MT as Resource” approach, implemented through the Integrate-Educate-Model strategy, provide a theoretical and practical foundation for incorporating MT into L2 education. Building upon these concepts, we can further conceptualise MT use in language learning through the lens of “MT as Augmented L2 Competence.” This model offers a visual representation of how MT can enhance learners’ language abilities, particularly in bridging the gap between receptive and productive skills. By viewing MT as a tool for augmenting competence rather than replacing language learning, we align with the translanguaging idea of fluid language practices and the “MT as Resource” approach. The Integrate-Educate-Model strategy can then be applied to help learners effectively utilise MT to expand their augmented competence zone, while simultaneously developing their own language skills. This integrated perspective not only justifies the use of MT in language learning but also provides a framework for understanding its role in enhancing overall L2 proficiency. The concept of MT as augmented L2 competence is illustrated in Figure 1, which provides a visual representation of how MT, and also GenAI such as ChatGPT, can enhance language learners’ abilities. 239 Atsushi Mizumoto Augmented competence with MT (AI) Productive competence Receptive competence Figure 1: The concept of MT as augmented L2 competence The figure demonstrates the relationship between receptive competence, productive competence, and the potential for augmented competence through MT use. Here is a breakdown of the key elements: • Receptive Competence: This is represented by the larger, outer oval. It refers to the ability to understand the target language (L2), which is typically more developed than productive skills. For most L2 English language learners, their capacity to comprehend English exceeds their ability to produce it. • Productive Competence: Shown as the smaller, inner oval, this represents the learner’s ability to actively use the language. It is generally more limited than receptive competence, which aligns with theories like Swain’s output hypothesis (1985), emphasising the importance of language production in second language acquisition. • Augmented Competence with MT (AI): This is depicted by the dark gray area extending beyond the productive competence oval. It illustrates how MT can bridge the gap between what learners can recognise as correct (receptive knowledge) and what they can produce on their own. Figure 1 suggests that MT can serve as a tool to augment learners’ competence, particularly in areas where they can recognise correctness by sight but struggle to produce it accurately. This augmentation is especially beneficial for more proficient learners, as supported by previous studies (Klimova et al. 2022, Ohashi 240 12 Embracing machine translation in L2 education: 2024). Higher proficiency learners tend to have a larger gap between their receptive and productive skills, providing more room for MT to assist in bridging this divide. Importantly, this model underscores that there remains a strong rationale for studying English (or any L2). The augmented competence provided by MT is built upon the foundation of the learner’s own language skills. Without developing one’s own receptive and productive competencies, the benefits of MT augmentation would be limited. Furthermore, as learners’ proficiency increases, they become better equipped to effectively utilise MT, maximising its potential as a learning tool. This conceptualisation of MT as augmented L2 competence aligns with the findings from systematic reviews (Jolley & Maimone 2022, Lee 2023) that highlight MT’s effectiveness when used appropriately, particularly for more advanced learners. It also supports the need for proper guidance and training in MT use, as the tool’s effectiveness is contingent upon the learner’s ability to critically evaluate and apply its output. In sum, the model presented here provides a framework for understanding how MT can be integrated into language learning processes. It emphasises that MT is not a replacement for language study, but rather a tool that can enhance and extend learners’ existing competencies, potentially accelerating their progress towards higher levels of language proficiency. 2.2 MT instruction for L2 learning Niño (2009) proposed four models of MT use in L2 education: a “bad model”, a “good model”, vocational applications (particularly in translation-related fields), and as a computer-assisted language learning (CALL) tool. Initially, MT was employed as a “bad model”, where students identified and corrected errors through post-editing, a process necessitated by the limited accuracy of early systems. In contrast, the “good model” involved using MT outputs as exemplars for students. These models, reflecting the evolution of MT technology and its pedagogical applications, illustrate a significant shift in focus. As MT technology has advanced, its primary use has transitioned to serving as a CALL tool, where it facilitates student engagement in solving language problems independently, as evidenced in recent studies (Lee 2020, Stapleton & Kin 2019, Tsai 2019). Lee (2023) recommends that teachers should provide guidelines for using MT and explicitly teach effective strategies to students prior to using it, which leads to enhancing student performance, as supported by O’Neill (2016). 241 Atsushi Mizumoto While many researchers discuss the need for explicit MT instruction in language classrooms, only a limited number of studies provide concrete examples (Cancino & Panes 2021, Chang et al. 2022, Mirzaeian 2021, O’Neill 2016, Ryu et al. 2022), all of which report positive results. Among these, Ryu et al. (2022) implemented a model called the Guided Use of MT (GUMT). The GUMT model activities were developed and implemented in an upper-elementary Korean as a foreign language course at a U.S. university. The GUMT model has five steps: • Instructional Session on MT Tools: As Activity 1, Students review strengths and weaknesses of popular MT resources like Google Translate and Naver Papago, discussing differences in speech styles and politeness levels. As Activity 2, students engage in pragmatic and grammar evaluation exercises, assessing the appropriateness of translations by considering contexts such as formality and speaker-listener relationships. • Practice with Writing Assignments: Throughout the semester, students apply learned concepts in five writing tasks, using MT to aid their translation from Korean (L2) to English (L1) and vice versa. • Reflection on MT Use: After each writing assignment, students reflect on their use of MT, assessing its effectiveness and making adjustments based on their experiences. • Instructor Feedback: Instructors provide written feedback on students’ drafts, focusing on grammatical and pragmatic accuracy, and highlight areas needing correction. • In-Class Review Sessions: Students refine their drafts in class based on instructors’ feedback, discussing and revising highlighted issues to improve their understanding and application of MT in writing. Ryu et al. (2022) reported that implementing the GUMT model significantly contributed to the development of MT use strategies and enhanced students’ confidence and self-assessed fluency in L2 writing, as evidenced by analyses of both preand post-surveys and student reflections. 2.3 Learners as metacognitive agents The role of learners as metacognitive agents in using MT and other language resources is crucial for effective language learning. Mizumoto (2023) proposed a 242 12 Embracing machine translation in L2 education: framework called Metacognitive Resource Use (MRU), which provides a comprehensive approach to understanding how learners can strategically utilise various language resources, including MT and GenAI tools like ChatGPT (Figure 2). Figure 2: Framework of metacognitive resource use (adapted from Mizumoto 2023). The MRU framework is grounded in metacognition theory and consists of two main components: metacognitive knowledge and metacognitive regulation. Metacognitive knowledge encompasses understanding of the person (selfknowledge), task (requirements and constraints), and strategy (available approaches). Metacognitive regulation involves the active management of this knowledge through planning, monitoring, and evaluating one’s use of resources through employing learning strategies. The framework visualises a range of resources that learners can utilise, including online dictionaries, corpus-based data-driven learning (DDL) tools like concordancers, web apps, search engines, GenAI systems like ChatGPT, MT tools, and grammar correction tools such as Grammarly. This reflects the real-world usage of tools by learners when engaging with language tasks. 243 Atsushi Mizumoto For instruction on metacognitive resource use, Mizumoto (2023) suggests adopting principles from learning strategy instruction (Chamot & Harris 2019). This approach involves enhancing learners’ metacognitive knowledge and imparting specific metacognitive and related cognitive strategies. The instructional model focuses on exposing students to a diverse array of tools and guiding them through the processes of planning, executing, evaluating, and adapting their strategic resource use. It is important to note that the GUMT model employed by Ryu et al. (2022) aligns well with this approach. The GUMT model does not rely on a single tool but incorporates various resources, including different MT platforms and reflective practices. Furthermore, the instruction model in Ryu et al. (2022) shares similarities with both DDL and strategy instruction approaches, demonstrating that these methodologies can be discussed within the same theoretical framework that Mizumoto (2023) advocates for integrating DDL and GenAI. This integrated perspective allows for a more comprehensive approach to language resource use. Instead of focusing solely on MT, learners should be encouraged to become metacognitive agents capable of effectively utilising a wide range of resources, including GenAI tools like ChatGPT. By adopting this approach, we can foster the development of autonomous learners who are adept at selecting and using the most appropriate tools for their language learning tasks. The MRU frameworknot only enhances learners’ awareness and use of MT but also extends to their conscious utilisation of GenAI tools. This holistic approach has the potential to cultivate truly autonomous learners who can navigate the complex landscape of language learning resources effectively. Moreover, the research findings accumulated in the field of MT should not be disregarded but rather re-examined and validated within this integrated and extended framework. By doing so, we can ensure that the valuable insights gained from MT research continue to inform and enhance our understanding of language learning processes in the era of diverse digital resources. In conclusion, by positioning learners as metacognitive agents within the MRU framework, we can create a more comprehensive and effective approach to language learning that embraces the full spectrum of available resources, from traditional tools to cutting-edge AI technologies. 2.4 Distinguishing MT from GenAI While this chapter has primarily focused on MT, it is important to acknowledge the emerging role of GenAI tools like ChatGPT in L2 education. Although both MT and GenAI can assist with language tasks, they differ fundamentally 244 12 Embracing machine translation in L2 education: in their capabilities and potential applications. Traditional MT systems are designed specifically for translation between languages, focusing on maintaining semantic equivalence while adapting to target language conventions. In contrast, GenAI systems can engage in open-ended language generation, including not only translation but also extending to explanation, summarisation, paraphrasing, and interactive dialogue. The differences between these technologies extend to their ability to handle context and maintain coherent interactions. While MT primarily processes text at the sentence or paragraph level, GenAI can maintain longer conversations, take broader context into account, and engage with complex prompts across multiple exchanges. This enhanced contextual awareness allows GenAI to serve not just as a tool for language conversion, but as an interactive partner in the language learning process. These fundamental differences lead to distinct educational applications. MT serves primarily as a translation tool, though it can be used pedagogically for language comparison and error analysis. GenAI, however, can potentially function as an interactive tutor, providing explanations, examples, practice exercises, and feedback on language use. This broader functionality suggests that while MT typically serves as a standalone tool for translation tasks, GenAI has the potential to integrate and potentially replace multiple language learning tools, including dictionaries, grammar checkers, and corpus resources. 3 Concluding remarks Looking ahead, the integration of MT, GenAI, and other AI-powered language tools in L2 education is likely to continue evolving rapidly. Future research should focus on developing and empirically testing comprehensive frameworks like the MRU, which embrace a wide range of digital resources beyond just MT. As the capabilities of GenAI continue to expand, there is a particular need to understand how these tools can complement or potentially transform traditional MT use in language learning contexts. There is a need for longitudinal studies to assess the long-term impact of these tools on language acquisition and learner autonomy, specifically examining how the interactive capabilities of GenAI might differ from traditional MT in supporting language development. Additionally, as AI language models become more sophisticated, practitioners and researchers will need to continually adapt their approaches, balancing the benefits of these tools with the core objectives of language learning. The field may see a shift towards more personalised, AI-assisted language learning experiences, necessitating new pedagogical strategies and ethical guidelines for their implementation 245 Atsushi Mizumoto in L2 learning and teaching. Understanding the distinct roles and potential synergies between MT and GenAI will be crucial for developing effective educational practices in this rapidly transforming field. References Cancino, Marco & Jaime Panes. 2021. The impact of Google Translate on L2 writing quality measures: Evidence from Chilean EFL high school learners. System 98. 102464. DOI: 10.1016/j.system.2021.102464. Chamot, Anna Uhl & Vee Harris (eds.). 2019. Learning strategy instruction in the language classroom: Issues and implementation. Bristol, UK: Multilingual Matters. Chang, Peichin, Pin-Ju Chen & Li-Ling Lai. 2022. Recursive editing with Google Translate: The impact on writing and error correction. Computer Assisted Language Learning. 1–26. DOI: 10.1080/09588221.2022.2147192. Chung, Eun Seon & Soojin Ahn. 2022. The effect of using machine translation on linguistic features in L2 writing across proficiency levels and text genres. Computer Assisted Language Learning 35(9). 2239–2264. DOI: 10.1080/09588221. 2020.1871029. Ducar, Cynthia & Deborah Houk Schocket. 2018. Machine translation and the L2 classroom: Pedagogical solutions for making peace with Google Translate. Foreign Language Annals 51(4). 779–795. DOI: 10.1111/flan.12366. Grieve, Averil, Amir Rouhshad, Elpida Petraki, Alan Bechaz & David Wei Dai. 2024. Nursing and midwifery students’ ethical views on the acceptability of using AI machine translation software to write university assignments: A deficitoriented or translanguaging perspective? Journal of English for Academic Purposes 70. 1475-1585, 101379. DOI: 10.1016/j.jeap.2024.101379. Hellmich, Emily A. 2023. Student use and instructor beliefs: Machine translation in language education. Language Learning & Technology 27(1). https ://hdl . handle.net/10125/73525. Jiang, Lianjiang, Rong Yu & Yi Zhao. 2024. Theoretical perspectives and factors influencing machine translation use in L2 writing: A scoping review. Journal of Second Language Writing 64. 101099. DOI: 10.1016/j.jslw.2024.101099. Jolley, Jason R. & Luciane Maimone. 2022. Thirty years of machine translation in language teaching and learning: A review of the literature. L2 Journal 14(1). 26–44. DOI: 10.5070/L214151760. 246 12 Embracing machine translation in L2 education: Klimova, Blanka, Marcel Pikhart, Alice Delorme Benites, Caroline Lehr & Christina Sanchez-Stockhammer. 2022. Neural machine translation in foreign language teaching and learning: A systematic review. Education and Information Technologies 28. 663–682. DOI: 10.1007/s10639-022-11194-2. Lee, Sangmin-Michelle. 2020. The impact of using machine translation on EFL students’ writing. Computer Assisted Language Learning 33(3). 1–19. DOI: 10. 1080/09588221.2018.1553186. Lee, Sangmin-Michelle. 2023. The effectiveness of machine translation in foreign language education: A systematic review and meta-analysis. Computer Assisted Language Learning 36(1-2). 103–125. DOI: 10.1080/09588221.2021.1901745. Lee, Sangmin-Michelle & Nayeon Kang. 2024. Effects of machine translation on L2 writing proficiency: The complexity, accuracy, lexical diversity, and fluency. Language Learning & Technology 28(1). 1–19. https://hdl.handle.net/10125/ 73585. Lo, Siowai. 2024. The effects of NMT as a de facto dictionary on vocabulary learning: A comparison of three look-up conditions. Computer Assisted Language Learning. 1–21. DOI: 10.1080/09588221.2024.2310285. Lo, Siowai. 2025. Neural machine translation in EFL classrooms: Learners’ vocabulary improvement, immediate vocabulary retention and delayed vocabulary retention. Computer Assisted Language Learning 38(3). 592–611. DOI: 10.1080/ 09588221.2023.2207603. Mirzaeian, Vahid R. 2021. The effect of editing techniques on machine translationinformed academic foreign language writing. The EUROCALL Review 29(2). 11. Mizumoto, Atsushi. 2023. Data-driven learning meets generative AI: Introducing the framework of metacognitive resource use. Applied Corpus Linguistics 3(3). 100074. DOI: 10.1016/j.acorp.2023.100074. Niño, Ana. 2009. Machine translation in foreign language learning: Language learners’ and tutors’ perceptions of its advantages and disadvantages. ReCALL 21(2). 241–258. DOI: 10.1017/S0958344009000172. Niño, Ana. 2020. Exploring the use of online machine translation for independent language learning. Research in Learning Technology 28. DOI: 10.25304/rlt.v28. 2402. O’Neill, Errol M. 2016. Measuring the impact of online translation on FL writing scores. IALLT Journal of Language Learning Technologies 46(2). 1–39. DOI: 10. 17161/iallt.v46i2.8560. Ohashi, Louise. 2024. Machine translation in language education: A systematic review of open access articles. Kenkyu Nenpou: The Annual Collection of Essays and Studies 70. 105–125. https://www.gakushuin.ac.jp/univ/let/top/ publication/KE_70/KE_70_008.pdf. 247 Name index Vieira, Lucas Nunes, 7, 126 Vinay, Jean-Paul, 97 Vygotsky, Lev, 33 Wadensjö, Cecilia, 214 Waldo, Jim, 63 Walker, Callum, iv Wang, Haifeng, 68 Wang, Longyue, 90 Wang, Lulu, 193 Wang, Rui, 70 Wang, Ziting, 211, 219, 220 Way, Andy, 68, 71, 80, 88, 90 Wei, Jason, 107 Wei, Li, 236 Welocalize, v Wiggers, Kyle, 109 Windhager, Florian, 8 Winters, James, 148 Wissen, Markus, 113 Woo, Jieyeon, 8 Wright, Sue Ellen, 68, 70 Yamada, Masaru, 56, 91, 93, 129, 132, 133 Yang, Jian, 148 Yang, Zhishen, 90 Zanettin, Federico, 68 Zhang, Biao, 90 Zhong, Linping, 192, 193 Zhou, Meng, 114 Zhu, Yilun, 148 Zhuang, Yan, 146, 148 Zimina, Maria, 155, 156 Zimina-Poirot, Maria, 149, 155, 157, 158 254 Teaching translation in the age of generative AI Since the launch of OpenAI’s ChatGPT in 2022, generative artificial intelligence (GenAI) has started reshaping what it means to work as a professional translator in an industry that is becoming increasingly automated. This prompts us to interrogate, once again, the role and agency of human translators in the translation process or, in other words, the intrinsically human value and values they add to it. A natural corollary is that GenAI forces us translator educators to (re-)interrogate what we do in our translation programmes. Whatever we may think or feel about GenAI, we owe it to our students to engage with it in our programmes. However, because GenAI is not just another tool in the translator’s toolkit, we must also to do so in a way that raises students’ awareness of some of the ethical and sustainability issues around it. This is what Teaching Translation in the Age of Generative AI: New Paradigm, New Learning aims to do. Articulated around three main parts, Part 1 explores the new skills and competences translator educators need to help their students develop in the age of GenAI. In Part 2, the focus shifts to the new knowledge (such as data literacy and prompting) that students in translation programmes need to engage with in the age of GenAI. Finally, Part 3 puts some flesh on the bones, as it reviews some of the new teaching approaches adopted by colleagues since the advent of GenAI. It does so by introducing the reader to a series of vignettes taken from a variety of translation-related disciplines and contexts. Throughout the entire edited volume, the ambition is to be as accessible as possible, so that this volume can be of help to as many of us in translation education as possible, as we all learn to negotiate the uncharted territory of GenAI.