The Project KIsu: AI-based searching and finding – easy, inclusive, and self-determined

THE PROJECT KISU: AI-BASED SEARCHING AND FINDING – EASY, INCLUSIVE, AND SELF-DETERMINED K. Altmeyer, M. Hladky, S. Malone, M. Platz, K. Reese, L. Schick, V. Wolf, Saarland University, Saarbrücken, Germany T. Gottsmann, M. Wiesner, fragFINN e.V., Berlin, Germany C. Plote, Open Search Foundation, Starnberg, Germany Abstract Many users, especially children, older people, or people with special needs, find it difficult to formulate precise search queries that lead to relevant results. This can lead to frustration, limited use of the internet, and ultimately, a feeling of digital exclusion. A lack of understanding of the search logic and an inappropriate interpretation of the search results harbor the risk of users consuming unreliable information or even being exposed to disinformation. In the KIsu project, we use an LLM that analyses and interprets the search terms entered. This AI model is trained to recognize and specify the user’s actual search intention from incomplete or imprecise queries. We are also developing workshops and information materials to promote the intelligent use of search engines and AI. KIsu is a cooperative project between Saarland University and fragFINN e.V. (as well as the German Research Center for Artificial Intelligence and Open Search Foundation). In this paper, the current status of the project is presented. INTRODUCTION Google dominates the search engine market, followed by Bing, which also provides the search results of most socalled alternative search engines and is gaining popularity thanks to the AI chatbot ChatGPT. These dominant providers do not disclose their algorithms and search indices, leading to restrictions in transparency, accessibility, data protection, and security, among other things. The consequences can be bias, information asymmetries, and datadriven discrimination (Bobic, Platz & Gütl, 2021 [1]; Galindo & Garcia-Marco, 2017 [2]). Although search engines play a significant role in our everyday lives, most people use them without knowing or questioning how they work. Effects such as a lack of transparency and immaturity are increasing massively due to generative AI language models. In addition, commercial providers have little interest in researching and offering low-barrier access. To enable all people – regardless of their age or gender, with or without disabilities – to use internet searches and search engines in an informed, self-confident, and self-determined way and thus facilitate their participation in the digitally pervaded world, we are developing learning and information materials and organizing workshops in the KIsu project. We want to support citizens in overcoming the black box effect of searches and analyzing and criticizing search results. Additionally, we aimed to specifically identify the challenges encountered by particular target groups during internet searches and to develop methods for supporting these groups both pedagogically and technologically. The learning materials on search engine literacy (e.g., Platz et al., 2023 [3]) and the Open Search Foundation (https://opensearchfoundation.org/en/children-and-internet-search/) served as a starting point. They are co-creatively adapted with various groups from civil society and further developed concerning the use of AI frontends and chatbots. We focus on children, young people and older people. All materials developed in the project are published as Open Educational Resources (OER). Design principles are derived and implemented for designing AI frontends for search engines that are orientated towards the common good based on the previously developed learning materials and explorative user workshops. This paper describes the objectives and steps of the KIsu project. Furthermore, the research design and first results of the project are presented. OBJECTIVES With KIsu, we want to strengthen the digital sovereignty and participation of people with special needs, children, and older people. We aim to enable as many user groups as possible to use search engines efficiently and inclusively. With the help of AI-supported language models, accessible search interfaces, and training, we create the conditions for using the internet as a source of information safely, confidently, and competently. In addition, our accompanying studies provide new insights into how generative AI models can be used to reduce existing barriers to internet use. In summary, KIsu pursues the following goals: •Promoting the competent and responsible use of Internet searches through developing and testing learning and information materials. •The identification of factors and specific needs in dealing with search engines and the derivation of implications for the further development of learning materials and for designing low-barrier AI search engine frontends. •Fine-tuning a needs-based AI language model and developing a model-agnostic data pipeline. •Provision of AI frontends for optimized search (technology maturity level 5). •Provision of the data collected in the project (the intervention and effectiveness study). https://doi.org/10.5281/zenodo.17229191 PROJECT STEPS The project duration is 19 months (01.06.202431.12.2025). The project consists of four steps: •Step 1 – Material and front-end development: learning, training, and information materials are designed co-creatively with various groups from civil society to promote the competent use of internet searches. The following group compositions and sizes were envisaged: 15 children, including children with various special educational needs (e.g., emotional-social development, mental development, hearing, physical and motor development, learning, vision, and language), 15 young people, including young people with various special educational needs, 15 older people, including people with various special educational needs. Care is taken to ensure a balanced gender composition. For the development of the first intuitive frontends for the search engines, a modular design is being sought that enables simple rule-based interactions and can be expanded later. Survey tools are selected and developed to identify factors and specific needs in dealing with search engines. A kick-off workshop with all project participants took place. Milestone: first version of materials and front end. •Step 2 – Material and front-end optimization: materials and frontends are tested with a controlled intervention study (n=45) and optimized based on the results. Milestone: Optimized version of materials and the front end. •Step 3 – Dissemination of the materials developed in the project and front-end development: A training concept and OER will be prepared in cooperation with various groups from civil society and published design principles for the design of low-barrier AI search engine front-ends are derived and implemented in a final application (training of a needs-based AI language model). The design principles for developing AI frontends to remove barriers to use and facilitate access to civic data will be made available. In addition, a conference will be held with all project participants. A multiplier network will be established. Milestone: training concept, OER, optimized version of the frontend, design principles, frontends. •Step 4 – Provision of data and publication of research results: The data collected will be processed and passed on to a research data repository. The research work and study results will be published for open access by the relevant specialist audience. Milestone: Open data, journal publications. RESEARCH DESIGN AND RESULTS An Action Design Science Research (ADSR) approach is pursued (Mullarkey & Hevner, 2019 [4]). Design Science Research (DSR) is a paradigm rooted in the philosophy of pragmatism. DSR involves problem-solving research to answer research questions related to human problems and produces valuable artefacts. ADSR centers on cocreative collaboration between scientists and users. The goals of the first phase (diagnostic phase) are to analyze the problem space and the solution space (here: the identification of factors and specific needs in dealing with search engines and the derivation of implications for the development of learning materials and for the design of low-barrier AI search engine front-ends) for research and practice and their relevance in mutual agreement between the researcher-user team. A mixed methods approach (e.g., Kuckartz, 2014 [5]) is pursued in which quantitative data collection methods such as questionnaires are combined with qualitative data collection methods such as interviews and observations in co-creative workshops. The sample comprises children, young people, and older people (see subsections below). The quantitative data is analyzed descriptively and inferentially, the qualitative data is analyzed using Design thinking methods, such as developing Personas (Uebernickel et al., 2015 [6]) combined with qualitative content analysis (Mayring, 2015 [7]), and the results are correlated. Then follows the design phase, in which the artefact is identified and conceptualized (here: learning materials and low-barrier AI search engine frontends). Design principles are (further) developed through several iterative cycles within the design phase. Collaborative activities with cocreative activities are essential here, as the researcher-user team aims to create artefacts that incorporate innovative ideas for solving the given problems. In the implementation phase, concepts are developed to use the artefact. An actual application offers the opportunity to evaluate the efficiency and effectiveness of the proposed design in practice. We are currently in the design phase (Step 2 of the project). We have already outlined detailed learning materials and selected specific tools to test the effectiveness of these. We will soon be testing these on an initial sample. At the same time, we are working on implementing the AI-supported search frontends and optimizing their functionalities based on user feedback. A central problem is that there has been very little sound research into the heterogeneous target groups and their behavior when searching the internet. We must, therefore, first create a solid empirical basis for further development steps. The project elements, co-creative workshops, a controlled intervention study, and the AI frontend are described below. Co-creative Workshops In line with Ind & Coates (2013 [8]), end-users are involved, which leads to more relevant and usable products and services while reducing risk. Participatory design is used to develop iterative prototypes to test user reactions. The workshops were conceptualized using Design thinking (e.g., Uebernickel et al., 2015 [6]). In the workshop the participants design their own digital assistant, that can help them to find what they search on the internet. In order not to tempt the workshop participants to reproduce existing solutions, but to become creative themselves, the word ‘digital assistant’ was used instead of ‘search engine’. The following keyquestions guided the workshops: •What are the features of your digital assistant? https://doi.org/10.5281/zenodo.17229191 oWhat does your digital assistant look like? oWhat should your digital assistant be able to do? •Input method: oHow do you want to tell your digital assistant what you are searching for? oWhat elements do you need on the screen to start your search? •Output method: oHow should your digital assistant present the search results? oHow should it tell you what it has found? oHow should the search results be presented? •What happens if you have (not) found what you were looking for? oHow do you tell your digital assistant? oHow does the digital assistant react to this? oWhat should the digital assistant do? The participants designed the interface of their assistant in small groups of 3-4 individuals, using small whiteboards and whiteboard markers as well as icons that could be stuck to the board (see Figure 1). Figure 1: Design of the digital assistant and input method by a child Four co-creative workshops were organized with the following user groups: •Children attending the 2nd grade in primary school (ca. 7 years old) o8 girls, 5 of whom speak German as a second language o7 boys, 3 with German as a second language and 1 with German as a foreign language •Mathematical interested children attending the 3rd or 4th grade in primary school (between 9 and10 years old) o5 girls o9 boys •Young people attending the 7th grade in grammar school (ca. 13 years old) o9 female o10 male •Older people (between 64 and 87 years old) o10 female o4 male During the workshops, qualitative interviews were performed with the participants. The workshops and interviews were videographed, and key scenes were transcribed. For analysis, personas are derived. Personas are descriptive models of users. They are archetypes with a set of properties of different but – concerning defined aspects –comparable persons (Uebernickel et al., 2015 [6]). Initial results indicate that children prefer having social and friendly interactions with their digital assistant, while trustworthiness is particularly important for older users. Speech input and output seem to be suitable across all user groups. Adolescents expressed a clear preference for tailored assistance, meaning the digital assistant should precisely match the complexity of the provided information to their specific needs and sensitively adjust its conversational tone – such as adopting a humorous style when searching for entertainment content. Controlled Intervention Study In the co-creative workshops, we observed that the group comprising slightly older children and teenagers demonstrated the highest level of prior knowledge regarding internet search strategies, use of information technology, and AI. Consequently, this group also provided the most substantial input for the co-creation of informational materials and the design of a suitable frontend for internet searches. Based on these insights, we will conduct an initial intervention study specifically targeting this age group, using tailored learning materials and a customized frontend designed for internet search activities. The study will address three research questions: 1. Does the use of specifically developed learning materials and the customized frontend lead to measurable improvements in aspects of search engine literacy and proficiency in AI-supported internet searches? 2. How do children evaluate the usability and usefulness of the developed frontend? 3. How do the designed informational materials impact children's interaction with the frontend during internet searches? The study will involve a minimum of 45 participants from grades 5 to 8 attending a Montessori school. The participants will work individually but will be organized into small groups for practical purposes. The study begins with all children completing pre-tests assessing their knowledge and attitudes toward internet searching and AI. Subsequently, children will be randomly assigned to one of two intervention groups: •Intervention Group 1 (IG 1) will first engage with the learning materials covering general information as well as input and output processes of AI-supported internet searches. They will then complete a structured internet search task using the newly developed frontend. Following this, participants will evaluate the frontend’s usability https://doi.org/10.5281/zenodo.17229191 and usefulness. Their interactions with the frontend will be recorded. Finally, the initial tests on search engines and AI will be repeated to measure learning outcomes and attitude changes. •Intervention Group 2 (IG 2) differs only in the sequence of tasks: children in this group will first work with and evaluate the frontend, followed by the study of informational materials. Overall, we expect both groups to benefit from our intervention by gaining essential knowledge related to critical aspects of AI-supported internet searches and, if present, correcting uncritical attitudes towards AI. Comparing the interactions between IG 1 and IG 2 will reveal the extent to which participants benefit from the learning materials during actual internet searches. It is hypothesized that participants in IG 1 will apply more of the principles covered in the materials compared to IG 2, leading them to perceive the developed frontend as more useful and usable. AI frontend A central aspect of our project is the development of innovative, low-barrier user frontends. The user interface should be able to recognize and understand the user’s questions, especially those of people with special needs. By integrating an LLM, the user query is analyzed, optimized, and converted into a suitable search query. This improved query is then forwarded to a search engine, e.g., FragFINN.de. There are also plans to include other (alternative) search engines (such as Ecosia). Despite the inherent opacity of large language models (LLMs), we would, like to use them specifically to dialogue with the user to better understand their actual informational needs and generate a more precise and relevant search request as part of our project. This process should enhance the understanding of users' actual informational needs of search results and accessibility and user-friendliness for all user groups. The ability to fine-tune LLMs for inclusive language or to adapt them to the unique search queries of children, older adults, or people with disabilities is central to us. Open-source models like Llama 3 or Mistral enable flexible adaptation through methods such as LoRA or QLoRA, while proprietary models like GPT-4-Turbo allow fine-tuning via API. Opensource models also have the advantage that they are more cost-effective, there is no direct dependency on individual companies, the dynamics of the models can be fully controlled, and data protection mechanisms are easier to implement and review. Through workshops and educational programs, we provide information about the data protection practices of search engines and the generative AI systems we use, which promote a more conscious and secure handling of personal data. The explainability and interpretability of the responses generated by LLMs are currently essential research topics. Although the presently available LLMs are not yet fully interpretable, we intend to actively follow the latest advances in this area and, where possible, integrate them into our project. This includes the evaluation and potential implementation of methods to increase the transparency and traceability of AI-supported processes. CONCLUSION The project KIsu aims to make digital information services accessible and understandable for everyone and thus strengthen digital participation in the long term. The project unfolds its impact through •the use of the training modules in training courses and train-the-trainer courses, •utilizing and making available the design principles and study results for the development of AI front-ends to reduce barriers to use and facilitate access to civic data, •building a multiplier network through training and OER, •raising awareness through public relations work. In the long term, we hope that AI will recognize users’ search intentions and proactively support them in specifying their queries, critically evaluating information, and making informed decisions. We also hope that the training concepts and technologies developed in our we recognize their potential to improve search queries significantly project will be integrated and used by educational institutions, social institutions, and advice centers to promote digital skills throughout society in the long term. ACKNOWLEDGMENTS We thank all the participants in the co-creative workshops and surveys, our student assistants and the German Federal Ministry for Family Affairs, Senior Citizens, Women, and Youth for funding the project, funding code 3924406K05. REFERENCES [1] Bobic, A., Platz, M. & Gütl, C. (2021). Towards Open Search Applications for the broader Community. In Proceedings of the 3rd International Symposium on Open Search Technology, 11-13 October 2021, CERN, Geneva, Switzerland. [2] Galindo, F. & Garcia-Marco, J. (2017). Freedom and the internet: empowering citizens and addressing the transparency gap in search engines. European journal of law and technology, 8(2),1–18. [3] Platz, M., Decker, A. & Plote, C.(2023). Förderung eines verständigen Umgangs mit Suchmaschinen im Elementarbereich. M. R. Textor & A. Bostelmann (Hrsg.), Das KitaHandbuch. [4] Mullarkey, M. T., & Hevner, A. R. (2019). An elaborated action design research process model. European Journal of Information Systems, 28(1), 6–20 [5]Kuckartz, U. (2014). Mixed methods: methodologie, Forschungsdesigns und Analyseverfahren. Springer-Verlag. [6]Mayring, P. (2015). Qualitative Inhaltsanalyse: Grundlagen und Techniken. Beltz Pädagogik [7] Ind, N., & Coates, N. (2013). The meanings of co‐creation. European business review, 25(1), 86-95. [8]Uebernickel, F., Brenner, W., Pukall, B., Naef, T., & Schindlholzer, B. (2015). Design Thinking: Das Handbuch. Frankfurter Allgemeine Buch. https://doi.org/10.5281/zenodo.17229191