Application of Generative AI in Health Systems: A Comprehensive Review of Innovations and Implications

Dissertation Title: "Application of Generative AI in Health Systems: A Comprehensive Review of Innovations and Implications" Master title: MSc. International Health Management Name: Mada Daniel Nkhoma Year: 2023/2024 ABSTRACT “The revolution of Generative Artificial Intelligence has unlocked an untapped U$1 trillion improvement potential in the healthcare industry.” Globally, there seems to be a paradigm shift regarding the future of health systems, especially with increased interest for digitisation, and the release of a Generative AI known as ChatGPT in 2022. Generative AI developments have materialised including those tailored for the health sector. The likelihood that GenAI together with other technological advancements will expand apace and become a vital requirement globally is certain. There is a growing certainty that people and machines may collaborate. This is particularly true with a transposition of Industry 4.0 (4IR) and Industry 5.0 (5IR). It seems Generative AI’s existence in today’s and future workplaces is unavoidable, but at what cost? Graphical Abstract 2 CONTENTS ABSTRACT 2 CONTENTS 3 ACKNOWLEDGEMENTS 5 DISSERTATION THESIS 7 INTRODUCTION 8 CHAPTER ONE – LITERATURE REVIEW I 14 1.1 Introduction 14 1.2 Theoretical Framework 14 1.3 Different Categories of Generative AI Models and their Applications 16 1.4 Benefits of integrating GenAI in health systems 19 1.5 Challenges and Ethical Implications of Integrating GenAI in health systems 24 1.6 Potential Future Research Directions and the Implications for Healthcare Organisations Considering the Adoption of GenAI 27 1.7 Summary 30 CHAPTER TWO – LITERATURE REVIEW II 31 2.1 Introduction 31 2.2 Generative AI Models: Use Cases, Applications and Benefits in Health Systems 32 2.3 Generative AI Tailored for Health Sector: Use Cases, Applications, Benefits and Challenges in Health Systems 35 2.4 Summary 43 CHAPTER THREE – METHODOLOGY 44 3.1 Introduction 44 3.2 Research Design 44 3.3 Population 44 3.4 Target population 45 3.5 Sampling Procedure 45 3.6 Sampling Technique 45 3.7 Sample Size 46 3.8 Inclusion Criteria 47 3.9 Exclusion Criteria 47 3.10 Data Collection Instruments 47 3.11 Data Collection Instrument 47 3.12 Pre-testing of Research Instruments 48 3.13 Reliability 48 3.14 Replicability 48 3.15 Validity 49 3 3.16 Data Collection Procedure 49 3.17 Ethical Considerations 49 3.18 Study limitations 50 3.19 Summary 50 CHAPTER FOUR – FINDINGS / ANALYSIS / DISCUSSION 51 4.1 FINDINGS 51 Section A: Demographic Data & Knowledge Assessment 52 Section B: Generative AI Applications and their benefits in Health Systems 55 Section C: Challenges and Ethical Implications of Generative AI in Health Systems 57 Section D: Future Direction of GenAI in Healthcare Management 58 4.1.1 Summary 60 4.2 ANALYSIS 61 4.2.1 Data Analysis Tools 61 4.2.2 Data Analysis Tool 61 4.2.3 Analysed Data 62 Section A: Demographic Data & Knowledge Assessment 62 Section B: Generative AI Applications and their benefits in Health Systems 63 Section C: Challenges and Ethical Implications of Generative AI in Health Systems 64 Section D: Future Direction of GenAI in Healthcare Management 64 4.2.4 Summary 65 4.3 DISCUSSION 66 4.3.1 Demographic Data 66 Knowledge of Generative AI 68 4.3.2 Generative AI Applications and their Benefits in Healthcare Systems 68 4.3.3 Challenges and Ethical Implications of Generative AI in Healthcare Management 69 4.3.4 Future Direction of GenAI in Healthcare Management 70 4.3.5 Summary 70 CONCLUDING REMARKS 72 BIBLIOGRAPHY 74 APPENDIX 0: Informed Consent 86 APPENDIX 1: Participant Questionnaire 88 APPENDIX 2: Industry 4.0 vs Industry 5.0 95 APPENDIX 3: Evolution of Generative AI 96 APPENDIX 4: Definition of Key Terms 97 4 ACKNOWLEDGEMENTS I would like to acknowledge the following four groups of people for supporting me throughout my academic journey, they have shaped the excellence of my scholarship. These are: Firstly, Grena from the Umodzi Wa aMalawi Community in Germany who helped me to explore various sources of reliable data during the brainstorming stage. Secondly, I extend my sincerest gratitude to my supervisor, Zoi Kairou. Her expertise, encouragement, and constructive feedback have been instrumental in shaping this dissertation and my growth as a scholar. Furthermore, I would like to express my deepest gratitude to my girlfriend, Esnart Rumando, for disturbing me with phone calls most of the time. I was furious, but it turned out to be a wonderful refreshing time, a time to re-energise, as her soft-sweet-voice and her never-ending-stories have been a constant source of strength, and I am profoundly grateful for her necessary disturbances every step of the way. Finally, I would like to extend my heartfelt appreciation to the diligent efforts and insightful grammatical checks and overall feedback provided by Sammuel Shakunle, Jessica Shakunle and Tayo Shakunle. Their valuable contributions have significantly enriched the quality and depth of this dissertation endeavour, for which I am sincerely grateful. 5 Statement of compliance with academic ethics and the avoidance of plagiarism I honestly declare that this dissertation is entirely my own work and none of its part has been copied from printed or electronic sources, translated from foreign sources and reproduced from essays of other researchers or students. Wherever I have been based on ideas or other people's texts I clearly declare it through the good use of references following academic ethics. (In the case that it is proved that part of the essay does not constitute an original work, but a copy of an already published essay or from another source, the student will be expelled permanently from the postgraduate program). Name and Surname (Capital letters): MADA DANIEL NKHOMA Date: 10/03/2024 6 DISSERTATION THESIS (leave this page empty) 7 INTRODUCTION “The revolution of Generative Artificial Intelligence has unlocked an untapped U$1 trillion improvement potential in the healthcare industry” (McKinsey & Company, 2023). Globally, there has been a paradigm shift regarding the future of health systems, especially with the release of a Generative AI (GenAI) known as ChatGPT in 2022. There is a growing certainty that people and machines may collaborate on specific administrative tasks (Weerarathna et., al. 2023). This is particularly true with the transposition of Industry 4.0, which began in 2010 and Industry 5.0, which started fruition in 2020 (Xu et al. 2021; Appendix 2). According to McKinsey & Company (2023), integrating human skills with Generative AI tools improves performance. As the application of technological models to management increases, health systems management is not spared, and these innovations are not only beneficial but also pose challenges and implications (Yamin, 2018). This chapter focuses on a brief introduction, background to the study, statement of the problem, research objectives, research questions, methodology, the significance of the study, delimitations of the study and the limitations of the study. Background to the Study Artificial Intelligence (AI) widens human abilities. This is the ideal for ensuring competitive health institutions. Three main types of AI include Artificial Narrow Intelligence (ANI), Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI) (Figure 1). ANI is designed to perform a specific task, and pre-defined parameters limit their abilities. For example, medical diagnostics (Kuusi & Heinonen, 2022). On the other hand, AGI also referred to as “Human-Level AI” aims at creating intelligent machines that can autonomously solve any intellectual task which a person can do, including generating novel or creative information (Goertzel, 2014). 8 Alan Turing designed “the Turing Test” in 1950 and investigated the potential of Human-Level AI. He is credited with being the first to propose the concept of thinking machines or Generative AI (GenAI). Subsequently, in 1960 Joseph Weizenbaum created the first Generative AI known as the “Eliza chatbot.” The Turing test is used as an important benchmark for the ongoing development of AGI. ASI is the third and final main type of AI. It is a type of artificial intelligence anticipated to exceed human-level intelligence in broad cognitive spectrums. ASI will have the capacity to perform tasks beyond human ability (Bostrom, 2014). This paper investigates the advantages of incorporating Artificial General Intelligence, specifically Generative AI into health systems while investigating the potential implications for workflow, processes and the health system. Generative AI utilises artificially intelligent systems to describe algorithms which can be used to generate new and original content like text, images, video, or audio as a response to prompts initiated by users (Varghese & Chapiro, 2023). The majority of GenAI is developed using neural networks which are designed to imitate the structures of the brain like Deep Learning and Large Language Models (LLMs). Notable GenAI applications are ChatGPT, Google Bard, Midjourney and DALL-E. According to Bitran (2023), Generative AI ensures healthcare management efficiency by streamlining the workflow processes and facilitating the creation of digital twins (AWS, 2024). 9 1.3 Different Categories of Generative AI Models and their Applications Generative AI models continue to advance, and their applications span a wide range of domains, from creative arts to practical problem-solving in healthcare, economics, and so forth. According to Lawton (2023), the choice of model often depends on the specific use cases, the requirements of tasks and the type of data involved (Miller, 2022). Generative AI models gained traction in the mid-2010s after the development of Variational Auto-Encoders (VAEs), Generative Adversarial Networks (GANs) and diffusion models (Lawton, 2023; Choudhary et al. 2022). Breakthroughs in generative neural network models like Transformers which can analyse large datasets to automatically create Large Language Models (LLMs) came about in 2017 (Lin, 2022). In 2022, Neural Radiance Fields (NeRFs), a technique for generating 3-dimensional (3D) content from 2-dimensional (2D) images was introduced (Appendix 3). 3-dimensional images are key in various simulations, media, gaming and the Internet of Things (IoT). In healthcare, NeRFs can be useful in medical imaging by facilitating the creation of comprehensive anatomical and physiological structures from 2-D scans like Magnetic Resonance Imaging (MRIs). That is, it can reconstruct realistic representations of body tissues and organs, giving doctors and clinic managers useful visual context (Corona-Figueroa et al. 2022). VAEs are often used for generating realistic images, text, speech or videos and for tasks requiring probabilistic representations (Singh & Ogunfunmi, 2021). GANs are more applicable when generating realistic data and in image processing (Goodfellow et al. 2020). The diffusion model, on the other hand, performs better use cases related to image synthesis, video generation, and molecule designing (Yang et al. 2023). Furthermore, Transformers can be applied in natural language processing, computer vision, speech synthesis, music generation, and multimodal applications such as visual question answering and commonsense reasoning (Lin, 2022). Transformer-based models, such as GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder Representations from Transformers) are a breakthrough well known for their ability to generate coherent and contextually relevant text, and were popularised by OpenAI and Google respectively (Pratim, 2023). Recently, Transformers have been used by DeepMind for AlphaStar (Zhang et al. 2023). 16 Other notable GenAI models include Autoregressive Models, Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) (Sherstinsky, 2020). Autoregressive models generate sequences one element at a time, conditioning on previous elements. According to Burant (2023), Autoregressive models are useful when forecasting future events based on historical data. For example, predicting patient outcomes and assessing the relationship between physical health and psychological well-being. RNNs and LSTMs are often used for generating data sequences, such as text or music. Besides, RNNs and LSTMs can be applied in natural language processing, creative content generation, speech-to-text transcription, and machine translation (Bandi et al. 2023; Burant, 2022; Sherstinsky, 2020). Table 1: Examples of GenAI and Its Applications What it Generates Description Application Example (s) Text Chatbots or AI writing tools generate new text as a response to a prompt from the user (e.g. an answer to a question, a summary, a paraphrase or a translation) ChatGPT, DeepL Translator, QuillBot Paraphraser, Scribbr Text Summarizer, Bard & Bing AI Code Natural language processing (e.g. English; Chinese), Output text in different programming languages (e.g. JavaScript). OpenAI Codex, GitHub Copilot Images LLMs can be used to generate images rather than text. These apps take a text-based prompt (e.g. “The Measles patient with symptoms”) and turn it DALL-E, Prisma, Midjourney, and Stable Diffusion 17 into an image. Some modify a user-submitted image. Videos These can create whole videos. The technology for this GenAI is still improving Synthesia, Gen-2, Make-a-Video, OpenAI’s SORA Audio GenAI may be used to generate therapeutic music and synthesised voices. Text-to-speech with a versatile AI voice generator MusicLM, MusicGen, MuseNet, Murf AI Other GenAI has potential in other areas, for example, in hard sciences (e.g. predicting protein structures) and in robotics (e.g. turning text prompts into actions done by robots). UniPi, AlphaFold Source: (Caulfield, 2023; OpenAI, 2024) According to Caulfield (2023), there are a range of GenAI applications which can be utilised in various fields including health management (Table 1). Similarly, other developments have revealed GenAI tools specifically designed to improve the efficiency of health systems such as Azure AI Health Bot (Bitran, 2023). In addition to that, Epic and Microsoft partnered to integrate AI into Electronic Health Records (EHR), allowing healthcare managers and care providers to improve productivity & patient communication with AI-enabled solutions (Epic, 2023). Moreover, the Mayo Clinic, the largest integrated medical group practice in the world has adopted Generative AI App Builder to enhance the capacity of its health systems (Landi, 2023). 18 1.4 Benefits of integrating GenAI in health systems According to Yu et al. (2023), GenAI can revolutionise how health data and information are handled, processed and managed. Moreover, a systematic review of 60 selected papers that assessed the utility of GenAI in healthcare with ChatGPT as a case study revealed that 85% (51 of 60) of papers reviewed cited beneficial applications like streamlining the workflow, cost-saving, improved documentation, personalised medicine, and improved healthcare literacy, while a staggering 97% (58 of 60) raised concerns (Sallam, 2023). On the other hand, administrative expenditure in the U.S.A., the largest and most complicated healthcare model in the world, accounts for 15 to 30% of healthcare spending (Health Affairs cited in Adner & Weinstein, 2023). As such, GenAI is poised to improve billing and claims processing as well as resource management and quality. McKinsey & Company (2023) note that GenAI is poised to improve the efficiency and effectiveness of healthcare management operations due to its ability to automate and summarise data, even big data, within a short period. This reduces turnaround time as health administrators focus on other important tasks. For instance, BioNTech, a biotechnology company, acquired InstaDeep, a type of GenAI, and developed an early-warning system for novel COVID-19 variants (Boston Consulting Group, 2023). Structural modelling of the SARS-CoV-2 protein integrated with InstaDeep’s Generative AI capabilities enables the system to forewarn and alert researchers, vaccine developers, health authorities, and policymakers (Boston Consulting Group, 2023). In addition, healthcare could be revolutionised by improving documentation accuracy and leveraging structured and unstructured data to create accessible clinical and administrative records, generating simulations aimed at educating employees and patients in critical areas and creating meeting or workshop summaries. Furthermore, Natural Language Processing (NLP) models can be used to understand and extract information from unstructured healthcare records. This facilitates efficient data management, enables trend analysis, and supports clinical and management decision-making (Avanade, 2024; Dash et al. 2019). According to Dash et al. (2019), an NLP-based algorithm known as “Linguamatics” utilises an interactive text mining algorithm (I2E). 19 I2E is capable of extracting and analysing vast amounts of information useful for both clinical and administrative problem-solving (Figure 3). The structured data generated by text mining could be integrated into data warehouses, databases or business intelligence dashboards and may be utilised for prescriptive, descriptive or predictive analytical decision-making (Linguamatics, 2023). Results obtained using this technique are tenfold faster than other similar tools and do not require expert knowledge for data interpretation (Dash et al. 2019; Linguamatics, 2023). Figure 3: An NLP-based GenAI System Used in Big Data Retention and Analysis A schematic representation of the working principle of an NLP-based AI system used in gigantic data retention and analysis in Linguamatics. Source: (Dash et al. 2019) Many health systems are grappling with high costs of care, staff shortages and employee burnout (Deloitte Center for Health Solutions Survey, 2023). To solve this, Value-Based Care is key. It aims at ensuring quality of care while promoting employee wellbeing, and personalised and affordable care. Text prompts into the GenAI like GPT-4 could draft standard Value-Based Care and carve out contracts based on market trends and characteristics (McKinsey & Company, 2023; Jayakumar et al. 2023). According to Deloitte Health Care Consumer Survey (2023), over half of the United States population (53%) is hopeful that Generative AI could improve access to health care; while 47% surveyed responded that GenAI has the potential to make healthcare more affordable. On the other hand, respondents who had already experienced Generative AI were more optimistic. 20 That is; 69% believe it could improve access to healthcare. And 63% indicated that GenAI has the potential to promote Value-Based Care. Such results are based on a nationally representative survey of 2,014 US adults (Deloitte Health Care Consumer Survey, 2023). Hence, GenAI may catalyse the implementation of Value-Based Care (Jayakumar et al. 2023). Moreover, it is believed that Artificial Intelligence can improve all kinds of processes within both healthcare operations and care delivery (Davenport & Kalakota, 2019). For example, the cost savings that AI could bring to the healthcare ecosystem is a pivotal driver for the implementation of AI applications, especially with the emergence of Generative AI which is capable of creating new and original content critical for quick decision-making and reduction of turnaround time (Avanade, 2024). It is estimated that integrating AI systems in healthcare management could cut annual healthcare costs in the United States by U$150 billion by 2026 (Bohr, & Memarzadeh, 2020). According to Bohr & Memarzade (2020), a big part of these reductions in costs stems from shifting the healthcare model from a reactive one to a more proactive approach, with an emphasis on health management rather than disease treatment. Furthermore, Robotic Process Automation (RPA) associated with GenAI performs structured (repetitive) digital tasks for administrative purposes. For example, those involving information systems. RPAs handle administrative tasks efficiently, as if they were a human user following a particular script or some rules. Compared to other types of AI they are inexpensive, easy to program with a high degree of transparency in their actions. Moreover, incorporating Robotic Process Automation (RPA) applications concurrently with GenAI systems in health systems could significantly reduce costs and improve administrative procedures (MSG Survey, 2023). These technologies are needed in healthcare because, for instance, the average nurse in the United States spends 25% of work time on regulatory and administrative activities which could otherwise be done by RPAs (Berg cited in Bohr & Memarzadeh, 2020). RPAs can be used for a myriad of applications in healthcare, including the processing of claims, clinical documentation, management of revenue cycle as well as medical records. 21 Notably, RPA does not involve robots, but rather, only computer programs on servers (Bohr, & Memarzadeh, 2020). Hence, it relies on a combination of business rules, workflow and “presentation layer” integration with health information systems to act like a semi-intelligent user of the systems. In healthcare systems, RPAs are used for repetitive tasks such as prior authorisation, updating patient data records and billing among other tasks. When combined with other technologies like OpenAI’s DALL-E, a Generative AI for image creation and recognition, RPAs can be used to retrieve data from, for instance, faxed images to input it into transactional systems (Hussain cited in Bohr, & Memarzadeh, 2020). This makes healthcare processes more efficient, effective and inexpensive. GenAI could foster technical operational efficiency as they can be used to draft Request For Proposals (RFPs) and requisitions, generate reports and Key Performance Indicators, draft vendor communications, and create purchase orders based on supply levels (McKinsey & Company, 2023). Therefore, since health facilities and departments have operated in silos for generations, GenAI has the potential to improve interoperability of healthcare systems. In health systems, GenAI may use unstructured purchasing and accounts payable data and with the aid of chatbots, can address common hospital employee information and human resource questions. Hence, it could improve employee experiences and may reduce time and money spent on hospital administrative costs (McKinsey & Company, 2023). GenAI promises beneficial impacts in Population Health Management (GenHealth.ai, 2023). Generative AI models like Pre-trained Transformers are spontaneous. The integration of predictive analytics may provide more accurate and timely identification of high-risk patients and at-risk populations, enhancing care coordination, reducing administrative burden and improving overall population health outcomes. (Yaraghi, 2024). In other developments, healthcare organisations have also experimented with chatbots for patient interaction, mental health promotion and wellness programs, including telehealth (Bohr & Memarzadeh, 2020). These NLP-based applications could be useful for simple transactions like scheduling meetings or appointments. 22 In addition, Doo et al. (2023) indicates that Generative Artificial Intelligence is reliable when determining population health outcomes and trends retrospectively, as such, may be helpful when analysing past public health challenges (Doo et al. 2023). However, as GenAI models are still developing, they are unable to predict current and future population health-related data. Similarly, GenAI could help to address disparities in healthcare as it promises to benefit the uninsured population (Dhar, Fera & Korenda, 2023). People who are not covered are more likely to use Generative AI when seeking health care like mental health support, to find a doctor, or to identify the most appropriate or closest care location (For instance, an emergency room or a doctor’s office). This is because vulnerable populations like the uninsured tend to be price-sensitive and may not be able to afford as little as a doctor consultation fee (Kaiser Family Foundation cited in Dhar, Fera & Korenda, 2023). Moreover, AI-powered tools could provide access to life-saving care swiftly and reliably, including health expertise to millions of vulnerable people in need (WeForum, 2024c). Despite that, critics argue that GenAI could exacerbate disparities in healthcare as high-income countries can invest in and upscale the relevant technologies rapidly (WeForum, 2024b; Khan et al. 2023; Omiye et al. 2023). To make matters worse, an average person in rural areas or low-income locations experiences intermittent access to electricity and an average person does not have access to internet connectivity, rendering GenAI applications harder to adopt inclusively. Compared to traditional AI systems, which are generally rule-based or rely on pre-defined datasets, GenAI models possess a unique ability to create new content that is original and unprogrammed. This can result in outputs that are similar to the user prompt in tone, style, or structure. Hence, if designed thoughtfully and developed responsibly, GenAI can amplify health managers' capabilities in various domains such as support for decision-making, knowledge retrieval, answering questions, language translation when supervising multicultural teams, and automatic reporting (Yu et al. 2023). 23 1.5 Challenges and Ethical Implications of Integrating GenAI in health systems While GenAI presents promising results, harmful outcomes in the management of health services can arise. These may include a tendency for LLMs to hallucinate. According to the International Business Machines Corporation (IBM), hallucination is a scenario whereby GenAI chatbots or computer vision tools perceive patterns or objects that are non-existent or incomprehensible by human observers, generating outputs that are nonsensical or altogether false (Ji et al. 2023). This leads to incorrect or misleading results generated by the AI models (Hatem, Simmons, & Thornton, 2023). Errors such as insufficient training data, incorrect AI model assumptions made, or biases in the data which is used to train the model contribute to the hallucination problem (Google Cloud, 2024; Alkaissi & McFarlane, 2023). Bang et al. (2023) find that GenAI, like ChatGPT, are overall 63.41% accurate in 10 different reasoning classifications under non-textual reasoning, logical reasoning, and commonsense reasoning. In this regard, It can be argued that Generative AI is an unreliable reasoner. Besides, GenAI has an issue of algorithmic bias (Oniani, 2023). According to Panch, Mattie & Atun (2019), algorithmic biases exacerbate disparities in healthcare. Therefore, it may be detrimental for healthcare managers to rely solely on GenAI for problem-solving and critical answers. Contrary to Dhar, Fera & Korenda's (2023) belief that GenAI could reduce disparities in healthcare, a study by Omiye et al. (2023) shows that Large Language Models could potentially perpetuate harmful practices by promoting debunked racist ideas (Figure 3). Besides, minority groups may be presented with distorted outcomes which might be the result of biases when mining big data used to inform the development of such models. Apart from that, the under-represented minorities as a consequence of racial biases in the development of datasets could be presented with poor predictions and unrepresentative results (Khan et al. 2023; Omiye et al. 2023). Technological biases have been proven in medical practices. For example, there is systemic and structural racism in German medicine, tools like Oximeters have been found to record misleading diagnoses for black patients in Germany and there is the potential of GenAI to exacerbate such disparities if due diligence is not done (Deutsche Welle, 2024). 24 Figure 4: Large Language Models Propagate Debunked Race-based Medicine For each question and each model, the rating represents the number of runs (out of 5 total runs) with disturbing race-based responses. Red indicates a higher number of disturbing and potentially fatal race-based responses. Source: (Omiye et al. 2023) Although Dash et al. (2019) and Linguamatics (2023) claim the vast array of benefits of NLP-based models in healthcare, a survey involving 500 users of the best five chatbots utilised in healthcare in the United States of America indicates that clients expressed deep concerns as regards discussing complex health conditions through the aid of chatbots or revealing confidential information. Besides, a lack of user-friendliness was also cited as a deterrent (MedNews, 2019). Apart from that, there are regulatory challenges associated with Generative AI (Hacker, Engel & Mauer, 2023). According to the proposed European Union Artificial Intelligence Act of 2021, a lack of a proper regulatory framework to guide the development and deployment of AI agents in the human market is a serious issue (European Parliament, 2023). As Artificial Intelligence technology is evolving very fast, the questions of trustworthiness are inevitable. 25 2.2 Generative AI Models: Use Cases, Applications and Benefits in Health Systems As the field of Artificial Intelligence continues to evolve, Generative AI models and their applications could transform how administrative and other health management processes are done (Miller, 2022). For example, from creating executive reports to practical problem-solving in healthcare, health economics and so forth (WeForum, 2024c). As Lawton (2023) and Miller (2022) note, the choice of model often depends on the specific use cases, the requirements of tasks and the type of data involved. Following this background, Variational Autoencoders (VAEs) can be applied in the management of health services, particularly in the domains of human resource management and the execution of administrative tasks. In this regard, VAEs can generate synthetic data that can resemble real-world data, notwithstanding patterns and distributions found in health service management processes (Giuffrè & Shung, 2023; Bandi et al. 2023). Such data can be valuable for training machine-learning models used in the management of human resources for health. Additionally, VAEs can be applied to analyse historical human resource data, for example, staffing levels, workloads, and performance appraisal metrics due to their ability to recognise patterns (Choudhary et al. 2022). Therefore, by learning the underlying patterns in this and related data, VAEs can assist in historical work-related assessments like performance reviews, training records, and public health project evaluations and outcomes. Besides, it could be able to predict future staffing needs, optimise resource allocation, and ensure adequate workforce availability. Moreover, VAEs are adept at detecting anomalies in data (Gonzalez et al. 2021). As such, in the context of health services administration, Variational Autoencoders can be efficiently utilised to identify unusual patterns or deviations in administrative tasks, alerting management to potential issues such as billing errors, data entry mistakes, or irregularities in process execution (Gonzalez et al. 2021). Moreover, since VAEs are trained on big data, they could model complex correlations between resource allocation decisions and outcomes in health service management or administration (Lawton, 2023). Hence, VAEs could provide insights related to resource allocation, whether it be financial resources, equipment distribution, human capital or individual employee assignments. 32 As regards Human Capital Management, VAEs could analyse the skill sets, qualifications, and individual training records of healthcare professionals (Gonzalez et al. 2021). This could help in workforce planning by identifying features representing specific skills. Hence, they could assist in matching health sector personnel to tasks. Additionally, VAEs could process data related to employee satisfaction surveys, feedback, and wellness programs. On the other hand, Generative Adversarial Networks (GANs) could be employed to create simulations for employee training activities (Lv, 2023). This may include simulated scenarios for hospital administrative tasks, patient interactions, or emergency response, allowing health personnel to practise and enhance their skills in a controlled virtual environment. According to Lv (2023), GANs are capable of generating diverse and realistic conversational data. This data could be used to train tailored healthcare chatbots and virtual assistants. These AI-driven tools could then assist in administrative tasks like documenting presentations, answering employee queries, or providing information about health services which are offered by a particular health unit. Not only that, GANs could also enhance facial recognition systems used for access control in healthcare facilities (Shahbakhsh & Hassapour, 2023). This is particularly relevant in managing human resources as it ensures secure access to restricted areas, monitors attendance and enhances the overall security of health units. Apart from that, GANs could augment datasets related to administration work by generating variations of available data (Wu, Stouffs, & Biljecki, 2022). This can be beneficial for tasks such as document classification, data entry, and workflow optimization. Besides, it could also be used for voice synthesis to create realistic voice models. In the management of health services, this can improve the quality of telephony services internally with employees or externally with stakeholders, for instance, automated appointment reminders. Notwithstanding, GANs could assist in generating realistic and informative reports for administrative purposes quickly and reliably, thereby reducing turnaround time, as they can be utilised to automate the creation of management reports, financial summaries, and other documents essential for a well functioning health system (Wu, Stouffs, & Biljecki, 2022). 33 Another notable GenAI model is Transformers, a type of Deep Learning model (Lv, 2023; Pratim, 2023)). As discussed in the preceding chapter, an advanced type of Transformers currently shifting the paradigm is called Generative Pre-trained Transformers (GPTs). Transformers have demonstrated great versatility in Natural Language Processing (NLP) and sequential data tasks (Lv, 2023). Due to that, Transformers could be used to build chatbots or virtual assistants capable of understanding and responding to natural language queries about administrative tasks. In the context of health management, this could include inquiries about upcoming schedules, policies, leave requests, or general information about health services or the workplace. For example, by understanding and processing natural language requests, Transformers may be able to interact with scheduling systems to find available slots, confirm appointments, and send reminders to patients and hospital staff (Nah et al. 2023). Furthermore, Transformers can assist in streamlining the onboarding process for new healthcare personnel by providing information about organisational policies, procedures, and training materials (Doo et al. 2023). Additionally, they can answer questions and guide employees through the onboarding process. Also, they may be utilised to create internal memos and reports, and can summarise lengthy administrative reports, contracts, policies, and guidelines (Nah et al. 2023). GPTs can also facilitate the generation of accurate billing codes and could assist with medical coding tasks in clinical administration (Luo et al. 2022; Newton, 2023). This could help healthcare administrators to quickly extract key information and make informed decisions spontaneously (Avanade, 2024). Similarly, transformers could also facilitate efficient retrieval of rare and old health information from large document repositories, for example, when needed by auditors or for reference (Luo et al. 2022; Newton, 2023). Not only that, but Transformers could also enhance email management by automatically categorising, summarising, and responding to routine emails (Lv, 2023). This can help healthcare administrators focus on critical tasks while ensuring that routine communication is handled efficiently. As far as human capital is concerned, Transformers could be utilised for resume screening and candidate matching during the recruitment process. For example, they can analyse resumes by identifying relevant skills and experiences while matching candidates to specific job requirements. This could streamline the hiring process. 34 In healthcare units with multilingual teams or which serve patients from a multicultural background, Transformers could provide real-time translation services, facilitating effective communication and collaboration among diverse staff members (Luo et al. 2022; Newton, 2023). Similarly, Transformers could be integrated into backend systems to automate routine administrative tasks, such as updating data, record keeping, and data reconciliation, improving the overall operational efficiency of health services administration (Doo et al. 2023). 2.3 Generative AI Tailored for Health Sector: Use Cases, Applications, Benefits and Challenges in Health Systems According to Waldron (2023), there are four broad areas where Artificial Intelligence and Generative AI are leveraged in the healthcare ecosystem: The first one is administrative as well as operational efficiency. Secondly, AI is leveraged in clinician and care team support. Furthermore, inpatient and consumer engagement, and finally advancing research and development strategies of resilient health systems (Waldron, 2023). The two notable AI tools which are challenging the status quo in healthcare are: (i) Azure AI Health Bot: Applications and Benefits To begin with, Azure AI Health Bot is a cloud-based conversational Artificial Intelligence service offered by Microsoft Azure, designed specifically for healthcare organisations (Hoorne, 2023). According to Hoorne (2023), it enables healthcare workers to build, deploy, and manage AI-powered chatbots and virtual health assistants to engage with both patients and colleagues. Azure Health Bot incorporates advanced natural language understanding capabilities, allowing it to interpret and understand user inputs in natural human language (Bitran, 2023). This enables hospital or clinic administrators to interact in real time with the bot using conversational language (Microsoft, 2024). Azure is profound in the management of health services because Healthcare organisations can easily customise the design features and behaviour of the chatbot to align with the specific needs of a health unit and preferences of the task at hand (Bitran, 2023). This could include defining conversational flows, designing user interfaces, and integrating various departmental workflows to ensure seamless communication across departments (Hoorne, 2023). 35 In addition, Azure Health Bot supports deployment across multiple channels, including websites, mobile apps, messaging platforms (like Microsoft Teams, Skype, and Facebook Messenger), and voice-enabled devices like Microsoft Cortana and Amazon Alexa (Moret-Tatay, Radawski, & Guariglia, 2022). It provides a centralised platform for scheduling appointments, managing inquiries and disseminating information. By reaching each responsible employee using the platform of his or her choice, Azure could ensure effective and timely communication of urgent administrative matters like low stocks of clinical sundries and supplies which need an urgent refill (Bitran, 2023; Microsoft, 2024). With Azure Health Bot, scalable and flexible services can be adapted to changing demands and the changing needs of the healthcare landscape (Hoorne, 2023). By automating routine administrative tasks, Azure Health Bot is capable of optimising resource allocation within healthcare organisations. This could free up hospital administrative and clinic staff to focus on more complex or specialised tasks, thereby improving operational efficiency. Automation could enable healthcare organisations to reduce costs associated with administrative tasks, such as switchboard call management, appointment scheduling, and staff training. This could enable organisations to achieve cost savings while maintaining service quality. Similarly, Azure Health Bot collects valuable data and insights from interactions with patients and users. It gathers information on frequently asked questions, service inquiries, user preferences, and feedback (Bitran, 2023). This information can be useful to health managers, policymakers and disease surveillance officers to safeguard population health. Moreover, Azure AI Health Bot can seamlessly integrate with existing as well as upscaled healthcare system infrastructure, electronic health records (EHRs), patient management systems, and health information communication channels (Microsoft, 2024). As it consolidates information from multiple sources, streamlines workflows, and enhances data interoperability across the organisation, a well-strengthened health system can be enhanced (Moret-Tatay, Radawski, & Guariglia, 2022). Besides, it prioritises compliance with healthcare regulations and industry standards, such as the Health Insurance Portability and Accountability Act (HIPAA) in the United States and the General Data Protection and Regulation (GDPR) in the European Union (Microsoft, 2024). 36 Through this compliance, Azure addresses ethical and other fears that users may have, as it ensures data security, privacy, and confidentiality, safeguarding sensitive information shared during interactions with patients and users. Challenges Associated with Azure AI Health Bot While Azure Health Bot has a wide range of benefits as it provides several built-in functionalities and pre-built templates tailored for the health sector, configuring and setting up Azure Health Bot may require advanced technical expertise and familiarity with Azure services (Microsoft, 2024). This can be very expensive in the long term. Besides, healthcare organisations without dedicated technical resources may find it challenging to deploy and manage the bot effectively. On the other hand, Integrating Azure Health Bot with existing healthcare systems, such as EHRs or health management systems, could be complex, and may require additional development effort as special developers are required (Microsoft, 2024). According to Microsoft (2024), Azure Health Bot offers a pay-as-you-go pricing model. However, costs can escalate based on usage and the scale of deployment (Microsoft, 2024). Hence, there is a high probability of incurring exorbitant charges. As such, healthcare units need to carefully monitor usage and consider potential cost implications, especially for large-scale deployments. Another Azure AI Bot challenge is the potential dependency on the Azure Ecosystem (Epic, 2023; Bitran, 2023). Azure Health Bot is closely integrated with the Microsoft Azure ecosystem. In this case, healthcare organisations already invested in alternative cloud platforms or infrastructure may face challenges in adopting Azure Health Bot due to dependencies on Azure-specific services and technologies (TrustRadius, 2024). This could create a monopoly capital for Microsoft, which has the potential to render health units vulnerable. Apart from that, information asymmetry between the technology giant, the actual Generative AI, and adopters raises concerns of misrepresentation, intended or unintended (Bathaee, 2018). 37 Although Microsoft (2024) claims compliance with HIPAA and GDPR, providing checks and balances to monitor its continued compliance could be a challenge due to the black-box nature of AI as well as the complexity of the big data that GenAI processes (Dash et al. 2019). Also, training and maintaining the Azure Health Bot requires an ongoing, almost unbreakable, commitment, as healthcare organisations need to allocate resources for training the bot and updating its knowledge base regularly (Microsoft, 2023). (ii) Generative AI App Builder: Applications and Benefits Generative AI Application Builder (Gen AI App Builder) necessitates the development, fast experimentation, and deployment of Generative AI applications in healthcare without the need for deep experience in AI. Notwithstanding, Gen AI App Builder is not a stand-alone GenAI application, rather it facilitates the building of novel and tailored GenAI applications. Examples include Amazon Web Services App Builder and Google Cloud App Builder. Gen AI App Builder could be utilised to build tailored GenAI applications for the management of health services. It has already started making strides as big healthcare centres like the Mayo Clinic have adopted the Google App Builder swiftly (Landi, 2023). It is capable of preprocessing data, model selection and training, deployment, as well as ongoing monitoring. In data preprocessing, Python libraries (like Pandas, and NumPy), SQL databases, and Extract, Transform, Load (ETL) can be used to gather and preprocess healthcare data, which may include Electronic Health Records (EHRs), staffing needs, and information about the overall health system (Healthcare Finance News, n.d.). Subsequently, during model development, Gen App Builder can help in selecting appropriate as well as tailored generative AI models based on the specific application requirements (such as GANs, VAEs, and autoregressive models) (Lawton, 2023; Choudhary et al. 2022). It can train the models using preprocessed data to learn patterns, generate synthetic data, or perform other generative tasks relevant to healthcare management. At this stage of tailored GenAI development, tools like TensorFlow, PyTorch, Keras, OpenAI's GPT, and Hugging Face's Transformers could be of significant relevance. 38 Not only that, Gen App Builder could also be useful during the deployment phase of the model. For instance, it can deploy trained models for health management as web services or as an Application Programming Interface (API) to make them accessible to healthcare managers and other professionals, also referred to as end-users (Healthcare Finance News, n.d.). Besides, Gen AI App Builders may facilitate compliance with data privacy regulations such as HIPAA or GDPR and could help in implementing security measures to protect sensitive healthcare data (Waldron, 2023). Cloud platforms (such as Google Cloud or Microsoft Azure), FastAPI, and Docker could assist in ensuring the effective deployment of the model (TrustRadius, 2024). Additionally, Gen AI App Builder could be vital in integrating Generative AI applications with existing health management systems, such as Electronic Health Record (EHR) systems or health information surveillance systems (Epic, 2023). In this way, a seamless data exchange and interoperability platform could be created to facilitate adoption and usability. For example, the App Builder could integrate GenAI with Fast Healthcare Interoperability Resources (FHIR), Health Level Seven International (HL-7), and Substitutable Medical Applications and Reusable Technologies (SMART on FHIR) allowing health managers to communicate, exchange knowledge and administer duties across the entire health system and with external stakeholders seamlessly (Lekkala, 2023). Lastly, the Gen App Builder may facilitate implementation, monitoring and logging mechanisms. This is pivotal when tracking model performance, usage patterns, and potential issues in real-time. And it has the potential to continuously update and retrain models as new data becomes available or as healthcare management requirements evolve. For example, the GenAI App Builder can integrate with Elasticsearch, Logstash Kibana (ELK stack) or Grafana to monitor health administrative processes in real time. Hence it can alert potential health system cyber security threats and abuse (ENISA, 2023). 39 Challenges associated with the Generative AI App Builder In spite of several applications and benefits that the GenAI App Builder offers, it has its drawbacks as well. For instance, it demands special developers with adequate knowledge of health systems. Over time, developers may become reliant on the features and capabilities provided by the generative AI app builder platform, thereby restricting their ability to integrate external libraries, frameworks, or advanced functionalities which may be more suitable. Furthermore, depending on the Generative AI App Builder platform chosen, GenAI developers may face vendor lock-in, making it difficult to migrate to alternative platforms or customise the application beyond the platform's offerings (Opara-Martins, Sahandi & Tian, 2014). Moreover, GenAI App Builders may involve subscription fees, usage-based pricing models, or additional charges for premium features (Microsoft, 2024; Google Research, 2024). This can contribute to ongoing costs for development and deployment, further straining the health system which already operates with scarce resources. Also, developers could face retraining challenges when using a specific Generative AI App Builder, particularly if they are unfamiliar with the platform's interface, workflow, or underlying technologies. This can impact quality and therefore the entire health ecosystem. For instance, a developer whose background is based on Amazon Web Services App Builder may find it challenging to work on Google Cloud App Builder. This can increase turnaround time, opportunity costs and unintended consequences. (iii) Other Notable GenAI Tailored for the Health Sector To begin with, GatorTron. It is a type of NLP model (Fries et al. 2022). Also referred to as a “large clinical language model” trained using over 90 billion words of text, including over 82 billion words of de-identified clinical text (Yang et al. 2022). It is by far the largest clinical LLM. According to the University of Florida News (2023), investigators trained supercomputers to generate health-related records based on a novel model, GatorTronGPT, that functions similarly to ChatGPT. Despite being expensive to train, GatorTron can easily adapt to new tasks, making it suitable for the ever-evolving healthcare landscape (Fries, 2022). 40 GatorTron excels in understanding and generating human-like text, hence is valuable for processing and interpreting unstructured data sources in healthcare, such as clinical notes, employee feedback, or biomedical literature. In addition, GatorTron can serve as a virtual assistant, providing real-time information retrieval, clinical decision support, and expert recommendations to clinic managers, healthcare providers and policymakers. Besides, they offer personalised responses to queries, helping health personnel to access relevant medical and logistical knowledge seamlessly (University of Florida News, 2023). GatorTron can also be integrated into patient-facing applications, chatbots, or educational platforms to deliver personalised health information, answer common questions, and provide self-care recommendations (Yang et al. 2022). This could improve patient engagement and health literacy apart from streamlining administrative tasks and enhancing documentation quality and completeness. Hence, the value and quality of care may improve overall. Figure 7: An Illustration of GatorTron; An NLP Model In Action Source: (Fries et al. 2022) 41 3.12 Pre-testing of Research Instruments A pre-test is when a questionnaire is tested on a statistically small sample of respondents before the actual study. Pre-testing offers the opportunity to see what questions work well, what questions sound strange, what questions can be eliminated and what needs to be added (Cormack, 2001). 3.13 Reliability According to Cormack (2001), reliability is the ability of the research study to produce similar results consistently when redone or performed by another researcher or after a certain period. There should be essentially little to no difference in results if the same methodology is applied to the same participants in the future as long as the same approach is used (Bryman, 2021). The results from the questionnaire administered as part of this study consisted of answers collected over a fortnight. Since the respondents were not given any new information or a different questionnaire to fill out throughout this time, it can be concluded that this measure is dependable. However, suppose the survey questionnaire was readministered a year later, it is possible that the participants' understanding of GenAI systems would have metamorphosed over time. Hence, their responses could be different. This is something that might be contested. And indicates that the reliability of this particular research technique cannot be sustained. 3.14 Replicability According to Bryman (2021), reliability is strongly linked to replicability. Replicability is when research techniques employed in one study can be reproduced to perform a different investigation of the same nature. Therefore, it could be hard for another researcher to reproduce the study if the first investigator does not explain the study procedure adequately. In this study, the research methodology, design and procedure have been described extensively in various sections of this dissertation. 48 The survey questionnaire used in this study has been attached as an appendix and is made accessible to other researchers who would want to conduct a similar investigation. Following this background, the findings of this research may be regarded as highly dependable and reproducible. 3.15 Validity Polit & Hungler (1999) define validity as “the ability of the research methodology to measure what it intended to measure.” There is a consensus among Polit & Hungler (1999), Cormack (2001), Gray & Grove (2020) and Bryman (2021) that a valid measurement is predominantly reliable. That is; if a test produces accurate results, they should be reproducible. Although a valid methodology is always reliable, a reliable methodology is not always valid. After this investigation, it was decided that the research methods employed to answer the primary research question (s) were effective and conclusive. 3.16 Data Collection Procedure After getting informed consent from the participants, data was collected anonymously within a fortnight using Google Forms. The collected data was kept securely in an online folder using access credentials managed by the Berlin School of Business and Innovation (BSBI), a university which complies with GDPR guidelines. 3.17 Ethical Considerations Permission was sought from BSBI’s Examinations and Assessment Department, a committee that ensures academic quality and ethical conduct. The purpose and benefits of the study were explained to the participants who gave informed consent. Privacy, anonymity, confidentiality and no coercion were maintained throughout the study. The participants were not forced to answer questions and were free to withdraw from the study at any stage without being victimised. 49 3.18 Study limitations The study was conducted for a short period as the investigator was following a pre-planned academic schedule. The convenience sampling method was used which has an element of bias. Moreover, the instrument which was employed to collect data was developed and used for the first time by the investigator who has no experience in research. As such, the instrument may not yield accurate and detailed information despite being pre-tested for validity and reliability. The primary investigation is limited to Germany as a geographic work area for the participants. Besides, not all key informants were reached, the language was a barrier since the investigator was not able to communicate with some key informants in German, the official language of the research setting. 3.19 Summary This chapter addressed the research methodology, research design, population and sampling. Furthermore, the data collection procedure and research instrument that was used when collecting data have been presented. On the other hand, issues regarding validity, replicability, reliability, and ethical considerations during the study were also addressed. 50 CHAPTER FOUR – FINDINGS / ANALYSIS / DISCUSSION 4.1 FINDINGS According to James Madson University (n.d.), the findings section of a research paper describes what the investigator found after analysing the data. The primary purpose of this section is to use the data collected to answer the research questions posed in the introductory chapter, even if the findings challenge the initial hypothesis. This chapter focuses on the presentation of findings, the analysis of data obtained from the research participants as well as the discussion of such findings. Data is presented in the form of frequency tables, bar graphs and pie charts. The study findings are systematically arranged into four sections in relation to the research objectives. The sections are as follows; Demographic Data & Knowledge Assessment, Generative AI Applications and their Benefits in Health Systems, Challenges and Ethical Implications of Generative AI in Healthcare Management, and the section about Future Direction of GenAI in Health Systems. 51 Section A: Demographic Data & Knowledge Assessment Figure 8: Age Representation of the Participants: N = 35 Participants, all responded. Table 3: Measures of Central Tendency for Age of Participants Mean 27 years Mode 26-35 years Median 26-35 years Range 17 Years Figure 9: Gender of Participants: N = 35, of whom all responded 52 Figure 10 Composition of Health Workers in the Study Figure 11: Knowledge Assessment: N = 35, all responded 53 Figure 12: Utilisation of GenAI Applications: N = 35, of whom all respond Table 4: What is Generative AI? N = 35, all responded Response Frequency Percentage (%) AI which can produce text, images, video or other material 17 48.6% AI which responds to human prompts and produces new and original content 22 62.9% AI which can generate electricity 1 2.9% I don't know 7 20% 54 Figure 13: Familiarity with Generative AI: N = 35, of whom 32 responded Section B: Generative AI Applications and their benefits in Health Systems Figure 14: Benefits of GenAI: N = 35, 28 of whom responded 55 Figure 15: GenAI in Healthcare: N = 35, all responded Figure 16: GenAI Use Case, and By Whom: N = 35, 34 of whom responded 56 Section C: Challenges and Ethical Implications of Generative AI in Health Systems Figure 17: Ethical Concerns of GenAI: N = 35, all responded Figure 18: Cost-effectiveness of GenAI: N = 35, all responded 57 Section C: Challenges and Ethical Implications of Generative AI in Health Systems In Figure 17 above, 35 participants were asked about their fears regarding the integration of GenAI in health systems. 15 participants (42.9%) indicated privacy concerns, 10 respondents (28.6%) pinpointed trust issues, and 25 participants (71.4%) highlighted that GenAI may lead to over-reliance and automation bias, which may compromise quality of care. However, 5 respondents (14.3%) did not have any fears; while 1 participant reported other fears. In the further comments section provided, the other fears were highlighted as “GenAI may lead to loss of job and quality issues.” Figure 18 above illustrates that 18 participants (51.4%) perceived incorporating GenAI in health management as costly. While 12 participants were not sure about the expenses. On the other hand, 5 participants (14.3%) perceived GenAI as cost-effective. Section D: Future Direction of GenAI in Healthcare Management Out of 35 respondents surveyed, 22 (62.9%) reported optimistic prospects of GenAI in healthcare management, 9 (25.7%) had no idea about how GenAI will impact health management in the future. 4 participants (11.4%) were pessimistic about the future benefits of GenAI in health management. Table 5 above shows that out of 15 respondents, 7 participants were aware of Azure AI Health Bot, 2 knew Generative AI App Builder, 6 knew about Med PaLM, another 6 demonstrated familiarity with BioGPT, 3 knew about ClinicalBERT while only 1 knew GatorTron. All representing 48.7%, 13.3%, 40%, 40%, 20%, and 6.7% of the respondents per each category respectively. In Figure 20, when asked how soon they expect to witness GenAI actively incorporated in the health systems, only 1 (3.2%) indicated within 1 year, 5 (16.1%) indicated 2 years, 9 (29%) expected GenAI in 3 years while the other 5 (16.1%) thought 4 years was realistic. However, 11 (35.5) participants representing the majority of answers were more pessimistic and expected to see GenAI being integrated into health systems in 5 or more years to come. 64 Figure 21 above shows that 5 respondents were sceptical about the future integration of GenAI in health systems, and 15 participants were positive that sooner or later GenAI will be integrated in health systems. Moreover, 14 participants were not sure. 4.2.4 Summary This section has highlighted the meaning of data analysis and different ways of analysing data in research, and how the collected data for this study was analysed. 65 4.3 DISCUSSION This section aims at interpreting and describing the significance of the research findings in light of what was already known about Generative AI in healthcare, and enables a thorough explanation of any new understanding or insights about the research topic after the findings and the analysis have been taken into consideration (Cormack, 2001). This section is written following a divide-and-conquer tactic of approaching a manuscript adapted from Şanlı, Erdem & Tefik (2013). The outcome of this dissertation has shed light on the possible application of GenAI in health systems by reviewing GenAI innovations and their implications. However, the findings ought to be interpreted carefully due to the limitations of this study. This dissertation investigated the application of GenAI in health systems. As discussed in the preceding chapters, GenAI has the potential to transform how the health sector operates. And several studies have been done about how GenAI will impact the healthcare industry. However, previous studies focused on the hospital and clinical processes with limited consideration for the health systems as a whole, hence, this study aims to fill that gap and make a contribution to further research. The third chapter is composed of a detailed explanation of the methodology and data collection procedure. Validity, reliability and replicability issues were also discussed, including the limitations of this dissertation. This chapter summarises key findings through the interpretations of the investigator regarding the significance of the study. Besides, the implications of this study are also taken into account through the lens of the investigator by focusing on the contribution of this study to what was already known. Following this chapter is the fifth chapter also referred to as “conclusion.” It presents a summary of the findings discussed and conclusions drawn. This is the final section of the dissertation. 4.3.1 Demographic Data In this study, 35 health workers in management or administrative roles participated, of whom 23 were female, representing 65.7% of respondents in the study. On the other hand, 12 were males representing 34.3% of participants in the study. These statistics confirm Statista's (2023) findings that in Germany, more females occupy leadership positions, especially in healthcare which accounts for 36.7% of all women in managerial roles. 66 Therefore, an open and inclusive acceptance and application of GenAI in health systems could ensure successful implementation. This further indicates that gender stereotypes in the technology field can only derail the effective adoption of Generative AI which is swiftly changing the definition of work, as both men and women are drivers of change, even, women arguably, have more to contribute as shown by the statistics above. The mean age of participants was 27 years, the mode was 26 - 35 years, and the median was also 26 - 35 years. The age range was 17 years. This shows that young adults are occupying leadership roles, and are shaping the perception and reality of the workplace. Statistisches Bundesamt (2024) concurs with these findings by stating that young persons are becoming highly qualified and are assuming leadership roles in the workplace. The number of qualified young people in Germany in creased by 10.8 per cent from 2002 to 2022 (Statistisches Bundesamt, 2024). In light of this, young people, when given adequate opportunities to thrive in bridging the technology acceptance gap in the health sector, more can be achieved. Similarly, young people are better qualified in technical pursuits, especially in the utilisation, development, and evaluation of Artificial Intelligence. 9 (25.7%) of health workers surveyed were in nursing administration, 6 (17.1%) were in health education, 5 (14.3%) were health managers, 5 (14.3%) were health researchers, 4 (11.4%) were in medical administration, 3 (8.6%) were in general public health, and 3 (8.6%) were clinic or hospital administrators. These results suggest that the nursing field has more health workers seconded by health education, then health management and health research. General public health and hospital or clinic administration were the least represented. Therefore, when considering the adoption of GenAI in health systems, it is vital to consider the allocation of technological resources and upskilling equitably. That is; according to need. 67 Knowledge of Generative AI Among the 35 health workers surveyed, 17 were familiar with Generative Artificial Intelligence while 18 were not. This indicates a knowledge gap regarding advancements in technology and the future of healthcare. Furthermore, when asked to name the Generative AI they know, a whopping 90.6% of respondents indicated familiarity with ChatGPT. That is; they were not aware or at least less familiar with other Generative AI applications like Murf AI. Besides, 80% of the participants reported that they have never used Generative AI applications at work; while 20% indicated that they have at least used the GenAI technology at work. These insights reveal a lack of awareness and technological literacy among healthcare workers. Moreover, knowledge of GenAI applications tailored for the health sector was also very limited, proving that even if the application of GenAI is to be adopted in health systems, its impacts might not be realised fully due to incompetent abilities or lack of readiness to realise it. 4.3.2 Generative AI Applications and their Benefits in Healthcare Systems 28 respondents familiar with GenAI applications were asked to mention the advantages of GenAI. 20 (71.4%) indicated that GenAI helps them to do tasks quickly and provides useful insights and answers. Such findings highlight the importance of GenAI in task optimisation and workflow improvement. This agrees with McKinsey & Company's (2023) and Bain & Company's (2023) findings that GenAI could improve the administrative processes and can optimise overall workflow. Nevertheless, only 13 (46.4%) respondents noted that GenAI improves task efficiency. This suggests that despite its beneficial impacts, users are still sceptical about the trustworthiness and accuracy of GenAI. Similarly, 14 (50%) reported that GenAI is helpful in problem-solving; implying that the other 50% did not regard GenAI as a good tool for decision-making. These findings refute the claims of Avanade (2024) and Dash et al. (2019) that GenAI could improve decision-making. Human beings are good at decision-making because of the ability to exercise emotions, moral judgement and empathy. As such, GenAI advancements can only be effective with humans-in-the-loop. 68 Despite that, 30 respondents representing 85.7% were optimistic about integrating GenAI in health systems. On the other hand, 5 respondents representing 14.3% were sceptical. This indicates that health workers are ready to embrace Generative AI with costs. For instance, out of 34 participants, 20 (58.8%) reported that GenAI should be an agenda for health managers. Whereas 14 (41.2%) participants indicated that health managers have nothing to do with GenAI, hence, GenAI should not be incorporated in health systems. These findings demonstrate a gap and a lack of consensus among health workers as regards whether or not GenAI should be applied in health systems. 4.3.3 Challenges and Ethical Implications of Generative AI in Healthcare Management Out of 35 respondents, 15 (42.9%) indicated privacy concerns, 10 (28.6%) pinpointed trust issues, 25 (71.4%) highlighted that GenAI may lead to over-reliance and automation bias. On the contrary, 5 respondents (14.3%) did not have any fears; while 1 participant reported other fears. In the further comments section of the questionnaire, the other fears were highlighted as “GenAI may lead to loss of job and quality issues.” These findings shed light on the serious ethical concerns, both held intrinsically at individual and corporate levels, associated with the incorporation of GenAI in health systems. This confirms the Deutsche Welle's (2024) report that the systemic design, the structural design and technological design of a health system can exacerbate health disparities. For example, tools like Oximeters have been seen calculating misleading diagnoses of black people because they were not designed as racially inclusive tools (Deutsche Welle, 2024). Therefore, the implementation of GenAI requires a thorough and methodical inclusivity in training data and racial representation. Apart from that, accountability issues, the fear of AI becoming super intelligent, and questions of morality and empathy were also concerns. Moreover, over half of the participants (51.4%) cited challenges related to affordability. As Microsoft (2024), one of the GenAI suppliers, notes, GenAI costs of developing, implementing and maintaining increases with the increase in usage and over a long period. In this way, health systems planning to or adopting GenAI are inevitably going to incur huge operational and related costs. Besides, health systems also risk being locked to the ecosystem of the GenAI supplier, which can be even more costly unless clear regulations are legally defined. 69 This vendor lock-in could be synonymous with holding health systems hostage. Furthermore, the asymmetric information between GenAI suppliers, GenAI itself, and users further worsen the fears associated with incorporating GenAI in the health system. 4.3.4 Future Direction of GenAI in Healthcare Management 22 (62.9%) respondents reported optimistic prospects of GenAI in health systems, and 9 (25.7%) had no idea about how GenAI will impact health management in the future. Meanwhile 4 participants (11.4%) were pessimistic about the future benefits of GenAI in health systems. Although fears and uncertainties persist, these findings highlight the readiness of the health workforce as regards the incorporation of GenAI in health systems. However, there is a lack of consistent consensus. Therefore, as regards the future, the adoption of GenAI in health systems might enable human-machine collaboration and accelerate the transition from Industry 4.0 (4IR) to Industry 5.0 (5IR). Notably, at the moment 5IR remains a utopia. Furthermore, 15 respondents demonstrated a lack of awareness of trending GenAI applications with the potential to impact the future of health systems. For example, only 7 participants knew Azure AI Health Bot, 2 knew Generative AI App Builder, 6 knew about Med PaLM, another 6 demonstrated familiarity with BioGPT, 3 knew about ClinicalBERT while only 1 knew GatorTron. This shows that GenAI may take health workers by storm if they are not intentional enough to get to know and assess their applicability in health systems. GenAI is inevitably transforming how management industries interact with employees. The question is, “Are health systems spared or is there a need for a proactive assessment of these transformative technologies?” 4.3.5 Summary This chapter has presented the findings of the study topic, "Application of Generative AI in Health Systems: A Comprehensive Review of Innovations and Implications.” The findings were presented in graphical, tabular and chart formats. 70 In addition, it has also highlighted how the collected data was analysed. Lastly, the findings have been discussed with reference to the reviewed literature and the analysed data from primary research. Similar or deviating findings from other studies and the observations and interpretations of the investigator were justified. 71 CONCLUDING REMARKS “The revolution of Generative Artificial Intelligence has unlocked an untapped U$1 trillion improvement potential in the healthcare industry.” Globally, there seems to be a paradigm shift regarding the future of health systems, and Germany is not spared, especially with the release of a Generative AI known as ChatGPT in 2022. Subsequently, significant GenAI developments have occurred including those tailored for the health sector. The likelihood that GenAI together with other technological advancements will expand apace and become a vital requirement globally is certain. There is a growing certainty that people and robots may collaborate on specific administrative tasks. This is particularly true with a transposition of Industry 4.0 (4IR), which began in 2010 and Industry 5.0 (5IR), which started fruition in 2020. Based on the investigator’s secondary and primary research, it seems like Generative AI’s existence in today’s and future workplaces is unavoidable. With reference to the Technology Acceptance Model which provides a theoretical framework for understanding and predicting users' acceptance and adoption of novel technologies, this study has reviewed literature related to GenAI. The review concurred with TAM that user acceptance of technology hinges on four factors, namely: perceived usefulness, perceived ease of use, social influence, and cognitive instrumental processes. These have been translated throughout this study as different categories of GenAI and their applicability in health systems, benefits of incorporating GenAI in health systems, challenges and ethical concerns of such innovation as well as present and future implications respectively. Generative AI innovations could catalyse the transition from 4IR to 5IR with the aid of human-machine collaboration, helping streamline health system processes, and facilitating the improvement of efficiency at the workplace. Whereas the implications have been identified as ethical concerns related to bias, morality, accountability, empathy and the fear of GenAI and associated technology becoming super intelligent. Moreover, implications about unsustainable costs, vendor lock-in and maintenance demands have been revealed. Notwithstanding, the asymmetric information among GenAI developers, Generative AI itself, and the end users. 72 The contributions of this study are: (1) To provide a basic understanding of generative artificial intelligence and reveal the potential value of GenAI in supporting healthcare systems, notwithstanding the drawbacks that come with such innovation. (2) To bring to attention and discuss the latest studies related to the application of GenAI in health systems, that provide a background highlighting the benefits and challenges of GenAI and potential future direction for further research (3) To foreshadow an image of a future health system and the potential redefinition of the role of the future health manager, including the skills that should be acquired and insist on the fact that there will always be a need for health managers, and human skills will possibly be more valuable in the future. As Generative AI continues to develop, the technology is becoming more sophisticated with time. There might be implications associated with such advancements. Therefore, the investigator provides 3 recommendations as follows: To start with, there is a need to ensure health workers, especially managers and administrators, have access to domain expertise or at least have prior knowledge or on-the-job training to make sense of GenAI advancements. Secondly, the creation of decisive and adaptable regulations and rules that must be applied to the datasets to generate useful insights without compromising the healthcare’s fundamental principles of “beneficence” and “nonmaleficence.” That is, rules and regulations to safeguard users from biassed AI solutions. Questions like, “Who will police the GenAI when a breach of GDPR, accountability or ethical misconduct is observed?” and “How will the issues of trust, morality and empathy be addressed?” should be considered. Finally, incorporating Generative AI in health systems may come with huge investment and opportunity costs both monetarily and ethically. As such, a thorough cost-benefit analysis must be done both qualitatively (how integrating GenAI will affect health workers and patients alike) and quantitatively (how integrating GenAI will impact costs and return on investment (ROI)) before, during and after the integration of GenAI into health systems. 73 Landi, H. (2023) Mayo Clinic, Google Cloud partner on Generative AI. Accessed from FierceHealthcare on 20.Jan.24 Lainjo, B. (2023) The Global Social Dynamics and Inequalities of Artificial Intelligence. Accessed from DOI:ResearchGate on 02.Feb.24 Lv, Z. (2023) Generative Artificial Intelligence in the Metaverse Era. Accessed from Cognitive Robotics on 02.Jan.24 Mascha, E. & Vetter, T. (2018) Significance, Errors, Power, and Sample Size: The Blocking and Tackling of Statistics. Accessed from PubMed on 12.Feb.24 McKinsey & Company (2023) Tackling Healthcare’s Biggest Burdens with Generative AI. [online] Accessed from McKinsey & Company on 11-Jan-24 Mahmud, M., Kaiser, M., McGinnity, T.M. et al. (2021) Deep Learning in Mining Biological Data. Accessed from Springer on 02.Feb.24 MedNews (2019) Consumers Don’t Fully Trust Healthcare Chatbots; Study finds. Health IT, Artificial Intelligence. Accessed from Health IT on 03. Feb.24 Miller, K. (2022) Deploying AI in Healthcare: Separating the Hype from the Helpful. Accessed from https://hai.stanford.edu/news/deploying-ai-healthcare-separating-hype-helpful on 12.Feb.24 McKinsey & Company (2022) Germany’s E-health Transformation Makes Uneven Progress. Accessed from McKinsey & Company on 17.Feb.24 Müller, J. (2020). Enabling Technologies for Industry 5.0—Results of a Workshop with Europe’s Technology Leaders. Directorate-General for Research and Innovation. Accessed from European Commission on 05.Jan.24 MSG Survey (2023) Over 80% of Companies in Germany Are Planning More Intensive Use of Generative AI. Accessed from MSG Group on 10.Jan.24 Meskó, B. & Topol, E. (2023) The Imperative for Regulatory Oversight of Generative AI in Healthcare. Accessed from Nature Digital Medicine Journal on 14.Feb.24 80 Moret-Tatay, C., Radawski, H. & Guariglia, C. (2022). Health Professionals' Experience Using an Azure Voice-Bot. Accessed from MDPI Journal 7.Feb.24 Microsoft (2024) Azure Health Bot: Pay-As-You Go. [online] Accessed from Microsoft Azure on 07.Feb.24 Nah, F., Zheng, R., Cai, J. et al. (2023) Generative AI and ChatGPT: Applications, Challenges, and AI-Human Collaboration, Journal of Information Technology Case and Application Research. Accessed from Information Technology on 13.Jan.24 Newton, W. (2023) What is BioGPT and What Does it Mean for Healthcare? Clinical Trials Arena. [online] Accessed from Clinical Trials Arena on 07.Feb.24 Omiye, J., Lester, J., Spichak, S. et al. (2023) Large Language Models propagate Race-based Medicine.Accessed from Digital Medicine on 03.Feb.24 Opara-Martins, J., Sahandi, R. & Tian, F. (2014) A Critical Review of Vendor Lock-in and Its Impact on Adoption of Cloud Computing. Bournemouth University, United Kingdom. Accessed from Bournemouth University on 07.Feb.24 Oniani, D., Hilsman, J., Peng, Y. et al (2023) Adopting and Expanding Ethical Principles for Generative Artificial Intelligence from Military to Healthcare. Accessed from Digital Medicine on 20.Jan.24 OpenAI (2024) SORA Text-To-Video: The Age of AI is Here. Accessed from https://www.youtube.com/watch?v=HK6y8DAPN_0 on 17.Feb.24 Polit, D. & Hungler, B. (1999) Nursing Research: Principles and Methods. ISBN-10: 0781715636 Pratim, P. (2023) ChatGPT: A Comprehensive Review on Background, Applications, Key Challenges, Bias, Ethics, Limitations and Future scope. Internet of Things and Cyber-Physical Systems. Accessed from IoT on 21.Jan.24 Peng, C. et al. (2023) A study of Generative Large Language Model for Medical Research and Healthcare. Accessed from Digital Medicine on 12.Feb.24 81 Panch, T., Mattie, H., & Atun, R. (2019). Artificial Intelligence & Algorithmic Bias: Implications for Health Systems. Accessed from Journal of Global Health on 20.Jan.2024 Qingxiong, M. & Liping, L. (2006). Technology Acceptance Model. Journal of Organisational and End User Computing. Accessed from ResearchGate on 6.Jan.24 Singh, A. & Ogunfunmi, T. (2021). An Overview of Variational Autoencoders for Source Separation, Finance, and Bio-Signal Applications. Accessed from MDPI Journal on 27.Jan.24 Sallam M. (2023). ChatGPT Utility in Healthcare Education. Systematic Review on the Promising Perspectives and Valid Concerns. Accessed from MDPI Journal on 19.Jan.24 Sherstinsky, A. (2020) Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) network. Accessed from Science Direct on 27.Jan.24 Shokrollah, Y., Yarmohammadtoosky, S., Nikahd, M. et al. (2023) A Comprehensive Review of Generative AI in Healthcare. Accessed from Cornell University on 30.Jan.24 Sage Research Methods (n.d.) Understanding Research. Accessed from Sage Publications on 17.Feb.24 Sezgin E. (2023) Artificial Intelligence in Healthcare: Complementing, not Replacing, Doctors and Healthcare Providers. Accessed from Sage Publications on 14.Feb.24 Shahbakhsh, M. & Hassapour, H. (2023) Empowering Face Recognition Methods using a GAN-based Single Image Super-Resolution Network. Accessed from Journal of Engineering on 24.Jan.24 Singh (2023) What is Research Design? Understanding Types of Research Design. [online] Accessed from Researcher Life on 13.Feb.24 Statistisches Bundesamt (2024) Quality Employment. [online] Accessed from Destatis.de on 20.Feb.24 Statista (2023) Share of Women in Leadership Positions in Germany in 2023, by Industry. [online] Accessed from Statista on 20.Feb.24 82 Sage Publications (n.d.) Literature Review and Focusing the Research. Accessed from Sage Publications on 04.Feb.24 Tellhed, U., Björklund, F. & Kallio-Strand, K. (2023) Men and Caring Women: Gender Stereotypes in Tech-Education. Sex Roles. Accessed from Springer on 20.Feb.24 Turchin, A., Masharsky, S. & Zitnik, M. (2023) Comparison of BERT Implementations for NLP of Narrative Medical Documents. Accessed from Science Direct on 07.Feb.24 Taherdoost, H. (2020) Different Types of Data Analysis; Data Analysis Methods and Techniques. International Journal of Academic Research in Management. Accessed from HAL Science on 19.Feb.24 TrustRadius (2024) Azure Health: Competitors and Alternatives. [online] Accessed from TrustRadius on 09.Feb.24 University of Florida News (2023) Medical AI tool from UF, NVIDIA gets Human Thumbs-up. [online] Accessed from University of Florida on 07.Feb.24 USC Libraries (2024) Research Guides: Organising Your Social Sciences Research Paper. Retrieved from http://libguides.usc.edu/writingguide on 1.Jan.24 Vogel-Heuser, B., & Hess, D. (2016). Guest editorial Industry 4.0–Prerequisites and Visions. IEEE Transactions on Automation Science and Engineering. Accessed from IEEE Xplore on 05.Jan.24 Verghese, J. & Chapiro, J. (2023) ChatGPT: The Transformative Influence of Generative AI on Science and Healthcare. Accessed from Journal of Hepatology on 14.Feb.24 WHO (2023) Digital Tools Positively Impact Health Workers’ Performance. Accessed from WHO on 6.Jan.24 Weerarathna, I., Raymond, D. & Luharia, A. (2023). Human-Robot Collaboration for Healthcare: A Narrative Review. Accessed from Cureus Springer Group on 3.Jan.24 83 WHO, OECD & WBG (2018) Delivering Quality Health Services: A Global Imperative for Universal Health Coverage. [online] Accessed from WHO, OECD & the World Bank Group Database on 4-Sept-23 WeForum (2024b) How Generative AI and Large Language Models Can Close the Gap between Data and Outcomes in Healthcare. [online] Accessed from World Economic Forum on 31.Jan.24 WeForum (2024c) 3 Ways Generative AI is Reshaping the Patient Experience. Health and Healthcare. [online] Accessed from World Economic Forum. On 03.Feb.24 WeForum (2023) Jobs of Tomorrow: Large Language Models and Jobs. A White Paper. [online] Accessed from World Economic Forum on 20.Feb.24 Wu, A., Stouffs, R. & Biljecki, F. (2022) Generative Adversarial Networks in the Built Environment: A Comprehensive Review of the Application of GANs Across Data Types and Scales. Accessed from Science Direct on 05.Feb.24 Waldron (2023) Generative AI in Healthcare: “Google Cloud’s Amy Waldron on the Tech Giant’s Health Ambitions” An Interview. [online] Accessed from Med-Tech Dive on 07.Feb.24 Wilkinson, D. & Dokter, D. (2023) The Researcher's Toolkit : The Complete Guide to Practitioner Research. 2nd ed. ISBN978-1032018102 Wits University (2023) Research Support Guide for Faculty and Students. Johannesburg. Accessed from Wits University Library. On 13.Feb.24 WeForum (2024) Patient-First Health with Generative AI: Reshaping the Care Experience. Accessed from World Economic Forum on 21.Fe.24 Xu, X., Lu, Y., Vogel-Heuser, B. & Wang, L. (2021) Industry 4.0 and Industry 5.0—Inception, Conception and Perception. Accessed from Science Direct on 3.Jan.24 Xiao, Y. & Watson, M. (2019). Guidance on Conducting a Systematic Literature Review. Journal of Planning Education and Research. Accessed from Sage Journals on 08.Feb.24 84 Yaraghi, N. (2024) Generative AI in Healthcare: Opportunities, Challenges, and Policy. [online] Accessed from Brookings.edu on 30.Jan.24 Yang et al. (2022) GatorTron: A Large Clinical Language Model. Accessed from Cornell University on 07.Feb.24 Yamin M. (2018). IT Applications in Healthcare Management: A Survey. International Journal of Information Technology. Accessed from. Journal of Information Technology on 11.Jan.24 Yang, L., Zhang, Z., Song, Y. et al. (2023) Diffusion Models: A Comprehensive Survey of Methods and Applications. Accessed from ACM Digital Library on 23.Jan.24 Yu, P., Xu, H., Hu, X., & Deng, C. (2023). Leveraging Generative AI and Large Language Models for Healthcare Integration. Accessed from MDPI Journal on 19.Jan.24 Zhang, E., Cheok, A., Pan, Z. et al. (2023) From Turing to Transformers: A Comprehensive Review and Tutorial on the Evolution and Applications of Generative Transformer Models. Accessed from MDPI Journal on 25.Jan.24 Şanlı, Ö., Erdem, S., & Tefik, T. (2013). How to Write a Discussion Section?. Turkish Journal of Urology. Accessed from Research and Practice on 02.Mar.24 85 Section A Demographic Data & Knowledge Assessment 1. Mark only one oval. 18 - 25 26 - 35 Over 35 Participant Questionnaire Introduction My name is Mada Daniel Nkhoma. I am a Berlin School of Business and Innovation student in collaboration with Università telematica internazionale Uninettuno headquartered in Rome, Italy. Itis part of the Erasmus+ and Erasmus Mundus programme. I am carrying out research on Generative Articial Intelligence (AI) in healthcare management. May you respond to the questions that follow at your convenience. The responses given will be anonymous and condential. Names will not be asked for on this form. It will take about 7 minutes to complete. If you would like to hear the results of this investigation, send me a message at Email: [email protected] Instructions 1. Answer as many questions as you can 2. Please do not write your names on any of these forms 3. Personal data including email addresses will not be collected or saved 4. Kindly answer the questions below, when necessary, include explanations in the spaces provided * Indicates required question Please provide the age range you belong to * APPENDIX 1 2. Check all that apply. 3. Mark only one oval. Health Services Manager Hospital or Clinic Administrator Researcher, Healthcare Other, Nursing Administration Other, Medical Administration Other, Public Health Human Resources for Health Other, Healthcare Education 4. Mark only one oval per row. Sex * Female Male Diverse Prefer not to say Your answer Your answer Which of the following fits well with your profession? Select an option closest to what you do * Do you know what Generative Artificial Intelligence is? * Yes No Not Sure Your answer Your answer Generative Al’s evolution For an advanced technology that’s considered relatively new, generative Al is deep-rooted in history and innovation. 1932 Georges Artsrouni invents a machine he reportedly called the “mechanical brain” to translate between languages ona mechanical computer encoded onto punch cards. 1966 MIT professor Joseph Weizenbaum creates the first chatbot, Eliza, which simulates conversations with a psychotherapist. 1980 Michael Toy and Glenn Wichman develop the Unix-based game Rogue, which uses procedural content generation to dynamically generate new game levels. 1986 Michael Irwin Jordan lays the foundation for the modern use of recurrent neural networks (RNNSs) with the publication of “Serial order: a parallel distributed processing approach.” 2000 University of Montreal researchers publish “A Neural Probabilistic Language Model,” which suggests a method to model language using feed-forward neural networks. 2011 Apple releases Siri, a voice-powered personal assistant that can generate responses and take actions in response to voice requests. 2013 Google researcher Tomas Mikolov and colleagues introduce word2vec to identify semantic relationships between words automatically. 2015 Stanford researchers publish work on diffusion models in the paper “Deep Unsupervised Learning using Noneq m Thermodynamics.” The technique provides a way to reverse-engineer the process of adding noise toa final image. 2018 Google researchers implement transformers into BERT, which is trained on more than 3.3 billion words and can automatically learn the relationship between words in sentences, paragraphs and even books to predict the meaning of text. It has 110 million parameters. Google DeepMind researchers develop AlphaFold for predicting protein structures, laying the foundation for generative Al applications in medical research, drug development and chemistry. OpenAl releases GPT (Generative Pre-trained Transformer). Trained on about 40 gigabytes of data and consisting of 117 million parameters, GPT paves the way for subsequent LLMs in content generation, chatbots and language translation. SYNTACTIC SiRUCTURE 2023 1957 Linguist Noam Chomsky publishes Syntactic Structures, which describes grammatical rules for parsing and generating natural language sentences. 1968 Computer science professor Terry Winograd creates SHRDLU, the first multimodal Al that can manipulate and reason out a world of blocks according to instructions froma user. 1985 Computer scientist and philosopher Judea Pearl introduces Bayesian networks causal analysis, which provides statistical techniques for representing uncertainty that leads to methods for generating content in a specific style, tone or length. 1989 Yann LeCun, Yoshua Bengio and Patrick Haffner demonstrate how convolutional neural networks (CNNs) can be used to recognize images. 2006 Data scientist Fei-Fei Li sets up the ImageNet database, which provides the foundation for visual object recognition. 2012 Alex Krizhevsky designs the AlexNet CNN architecture, pioneering a new way of automatically training neural networks that take advantage of recent GPU advances. 2014 Research scientist lan Goodfellow develops generative adversarial networks (GANs), which pit two neural networks against each other to generate increasingly realistic content. Diederik Kingma and Max Welling introduce variational autoencoders to generate images, videos and text. 2017 Google researchers develop the concept of transformers in the seminal paper “attention is all you need,” inspiring subsequent research into tools that could automatically parse unlabeled text into large language models (LLMs). 2021 Openal introduces Dale, which can generate images from text prompts. The name is a combination of WALL-E, the name of a fictional robot, and the artist Salvador Dali. 2022 Researchers from Runway Research, Stability Al and CompVis LMU release Stable Diffusion as open source code that can automatically generate image content from a text prompt. Opendl releases ChatGPT in November to provide a chat-based interface to its GPT 3.5 LLM. It attracts over 100 million users within two months, representing the fastest ever consumer adoption of a service. Getty Images and a group of artists separately sue several companies that implemented Stable Diffusion for copyright infringement. Microsoft integrates a version of ChatGPT into its Bing search engine. Google quickly follows with plans to release the Bard chat service based on its Lamda engine. And the controversy over detecting Al-generated content heats up. com rerarocr Aun ncenven TechTarget APPENDIX 3