Trust and overreliance on ChatGPT and its implications on critical thinking: an exploration in health policy

You’re listening to “Trust and overreliance on ChatGPT and its implications on critical thinking: an exploration in health policy,” by C.A. Rosas-Jiménez and colleagues. Published in 2026.

Abstract.

International organizations have provided guidelines for the use of AI, encouraging its responsible development and promoting human-centered, trust-based use. If governments or international organizations wish to encourage, discourage or even prohibit the use of ChatGPT or similar generative large language models, thorough ethical analyses are needed to inform decisions and recommendations. Rooted in the Technology Acceptance Model (TAM), this narrative review focused on the implications of trust and overreliance on ChatGPT and critical thinking and explored some of the implications of using this AI tool in health policy. Although trust in ChatGPT is not the same as trust in people, the use of ChatGPT is based on trust. However, blind trust, i.e. trust that is not aware of the limitations, drawbacks and challenges of ChatGPT, can lead to overreliance.

We suggested that overreliance on ChatGPT can come from two sources: First, from its excessive use. Second, from disregard or ignorance that there are certain limitations to using ChatGPT. We proposed six requirements that could contribute to optimal reliance on ChatGPT in health policy to avoid the loss of critical thinking skills and even improve those skills. The application of this model in health policy could be translated into requirements for future health policymakers using ChatGPT.

The World Health Organization in its report about Ethics and Governance of Artificial Intelligence (AI) for Health stated that the “use of AI for health is still new and often untested, and policymakers and regulators must consider numerous ethical, legal and human rights issues”. In line with these guidelines, the European Parliament recently launched the European Artificial Intelligence Act to promote the responsible development and use of artificial intelligence in the European Union (EU), which has been described as “the world’s first comprehensive legal framework dedicated to AI”. Related to article No. 1 of the Artificial Intelligence Act that encourages trustworthy and human-centered AI warranting fundamental rights, safety and health, the present narrative review focuses on the use of ChatGPT1, exploring some of the implications of using this AI tool in health policymaking.

If governments or international organizations wish to encourage, discourage, or even prohibit ChatGPT usage, evidence of what ChatGPT can do and thorough ethical analyses are needed to inform decisions and recommendations.

Correspondence:

Carlos Alberto Rosas-Jiménez

the email address

1Mary Heersink School of Global

Health and Social Medicine,

Canada

2Department of Health Research

Canada

3Centre for Human Rights and

Restorative Justice, McMaster

Users’ critical thinking could be affected by ChatGPT use. For example, a study of South Korean university students who used ChatGPT showed a decrease in critical thinking and a systematic review that analyzed 33 studies of limitations associated with ChatGPT found constraints in problem solving and critical thinking in 22% of cases. Losing critical thinking skills isn’t an isolated event, it has been associated with reduced creativity and independent thinking, as well as increased laziness and plagiarism.

The term critical thinking can be traced to John Dewey’s definition as “active, persistent, and careful consideration of any belief or supposed form of knowledge in the light of the grounds that supports it, and the further conclusions which it tends”. In the 1940s, a seminal study was published describing an experiment designed to develop critical thinking. Ennis stated that critical thinking is accurately assessing statements and that it can be broken down into three dimensions that can be distinguished analytically: a logical, a criterial, and a pragmatic dimension. Since then, there has been abundant research on critical thinking. Sanders & Moulenbelt made a list of definitions of critical thinking in chronological order, finding that a definition cannot be agreed upon.

The ability to think critically has been identified as a key skill, for example, for nursing and medical students. Some research on critical thinking includes exploring the elements that contribute to creating an educational environment that fosters critical thinking among university students, primary students, and teachers, as well as potential drawbacks of devoting extra time and resources to teaching general critical thinking. Besides, the importance for students to develop individual and social critical thinking when using AI has also been highlighted. Most critical thinking assessments measure the ability to analyze arguments, reason inductively and deductively, and reason quantitatively.

There have also been detractors of critical thinking. For example, between 1995 and 1997, there were some articles that oppose critical thinking’s inclusion in American postsecondary composition courses. In response to some of these criticisms, Benesch promoted the concept of “dialogical critical thinking”. Dialogic critical thinking has been defined “as a form of dialogical discourse in which the taken-for-granted assumptions and presuppositions that lie behind argumentation are uncovered, examined, and debated”. Although Benesch’s arguments had a more social and political connotation, she promoted the idea of dialogical thinking, which in conjunction with the concept of dialogue-based critical thinking are very appropriate nuances of critical thinking related to the use of chatbots, such as ChatGPT.

In this review we focus on the need to preserve and promote critical thinking skills when using ChatGPT and emphasize the importance of dialogic critical thinking. As it will be explained below, Users have two main opportunities to develop their critical thinking skills while they have a dialogic relationship with ChatGPT: first, in the generation of prompts; and second, in the analysis of the responses given by the chatbot.

Regarding the use of AI, Bitzenbauer recently presented examples of how to use ChatGPT in the classroom to develop students’ critical thinking skills, as long as students engage in a detailed verification process of the information provided by the teacher. Very often, if not always, the data that ChatGPT gives in its answers needs to be double-checked and corroborated with peer-reviewed articles or reputable sources. Therefore, educators should stress on the importance of critical thinking to help students use AI tools with a healthy dose of skepticism and integrating elements such as iterative learning, melioration and ethical reasoning offers educators and policymakers the opportunity to improve critical thinking in a setting where AI is a key player.

Since its launch, ChatGPT has rapidly grown in importance around the world. Its usage surpassed the previous record set by TikTok, reaching 100 million users just two months after its release. In addition, a recent study showed that ChatGPT was much better than Gemini and Claude on several metrics, with high levels of robustness, accuracy, recall and precision, demonstrating its advanced ability to handle complex reasoning situations. In light of the previous background, the aim of this narrative review is to explore the extent to which trust and overreliance on ChatGPT can influence critical thinking, focusing on health policymaking as a contextual example where ChatGPT can be useful.

To achieve this goal, we reviewed the literature related to the use of ChatGPT in health policy, giving priority to the most recent articles. The literature review was followed by a critical analysis. It is vital to note that most of the articles about the challenges, limitations, disadvantages, and ethical concerns of using ChatGPT found in the literature are commentaries, editorials, letters to the editor, and not peer-reviewed research articles. Many of the research articles were found in archive repositories (i.e., ​h​t​t​p​s​:​/​/​ a​r​x​i​v​.​o​r​g​/​) and are not peer-reviewed, which might be due to the recent emergence of ChatGPT. Therefore, there is a lack of research that can provide practical evidence of these challenges, limitations, drawbacks, and ethical concerns. It means that ChatGPT is being used without much evidence of the impact its use can have.

2 Exploring the impact of ChatGPT as an AI tool

The idea of creating intelligent machines has been present in human thought for centuries, with early references found in ancient myths and scriptures describing intelligent creatures forged by humans; for example, Talos, the “bronze giant” in Greek mythology, was created by Hephaestus, the Greek god of invention and technology. A detailed history of AI can be found in Ertel or much more recently in Anurag.

A milestone in artificial intelligence can be set in 1920, when Karel Capek’s Czech science fiction play Rossum’s Universal Robots first used the word “robot”; since then, there have been several developments to the point until the term “artificial intelligence” itself was introduced by John McCarthy in 1955. Later on, several definitions of AI have been given and also been analyzed from different perspectives, some of them analyzing the concept from what “artificial” and “intelligence” mean. In fact, Wang stated that there is no correct definition of AI because each definition has its own theoretical and practical values.

According to Yubing et al., the more common definition is: “artificial intelligence is the theory, method, technology and application system that uses digital computer or digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to achieve the best results”. Nevertheless, as Wang explains, the current field of AI is a mixture of several areas of research, each with its own goals, methods, and applicable situations. A recent survey of 23,882 people in 21 countries (> 1000/country) and in 12 languages, considered to represent a majority of the world’s population, found that 73% of the respondents felt they understand what AI means, with Indonesians having the highest self-reported knowledge—80% either agreeing or strongly agreeing that they know what it means.

The study also found that global attitudes toward AI are somewhat positive, with half of those surveyed stating that they feel somewhat (34%) or very (16%) positive about AI and19% of American respondents felt quite negative. Recent studies have also shown that some people are afraid of AI.

AI has been categorized in many ways. One of these categories is Natural Language Processing (NLP) systems to which ChatGPT belongs. The Chat Generative Pre-Trained Transformer (ChatGPT) is an intelligent chat robot able to give a detailed answer in accordance with an instruction in a prompt. The core techniques used by ChatGPT, such as NLP, Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) models, Transformers, and Reinforcement Learning from Human Feedback (RLHF), have revolutionized how machines comprehend and interact with human language. ChatGPT, like other large language models (LLM), is an NLP system trained on extensive data to respond to queries in a way that feels human. In fact, LLMs can perform both inductive and deductive tasks as humans can do.

According to Darwish, a chatbot is a program that emulates human communication and when this chatbot uses AI technologies, it is called an AI chatbot e.g., ChatGPT. Using conversational AI techniques and learning algorithms to improve their performance, virtual assistants engage in dialog, e.g., Siri (Apple) or Alexa (Amazon). Prompts are fragments or natural language-based sentences and by designing and refining the initial input, Prompt engineering plays a key role in shaping the responses of AI systems.

ChatGPT’s history, current status, and potential development for the future have been summarized, as well as the technical path from GPT-1 to GPT-4. There are several studies about prompt engineering with ChatGPT, its use in creating public policy, health policy analysis, education, healthcare and medicine, healthcare research, healthcare education and research and telemedicine, among several applications. In PubMed alone, > 7500 articles use the term ChatGPT as of January 2025. ChatGPT provides new capabilities such as human-like interaction and assistance, educational support and skill development, task automation and workflow improvement, content creation and ideation, and information retrieval and application.

ChatGPT has also been shown to play a positive role in improving students’ motivation and the independence and autonomy of the self-taught learner can increase with the use of ChatGPT. Some details about the use of ChatGPT identified in the survey by the Schwartz Reisman Institute for Technology and Society are in Table 1.

The reflection of this narrative review is rooted on the Technology Acceptance Model (TAM), proposed by Fred Davis, which is one of the most prominent frameworks to understand user acceptance and use of technology and has become a dominant model in the study of factors that affect user acceptance of technology. Derived from the Theory of Planned Behavior and the psychology-based Theory of Reasoned Action, the TAM plays a mediating role between perceived usability and usefulness. According to TAM, perceived usefulness is also influenced by perceived ease of use, because the easier the system is to use, the more useful it can be. Therefore, the TAM can provide a framework for developing the concept of trust in ChatGPT.

The use of ChatGPT has recently been better understood through the TAM, which has proved very useful in understanding, for example, PhD students’ intentions to accept written ChatGPT and what factors might influence their intentions to accept written ChatGPT. TAM, one of the most influential and parsimonious theories for prediction of user acceptance, is influenced by trust. Applying the TAM, it has been shown how ChatGPT trust considerably facilitates the relationship between ChatGPT self-efficacy and actual ChatGPT use among Pakistani university students, highlighting the role of trust as a driver to embrace new technologies such as ChatGPT. Other authors have contributed to using ChatGPT by emphasising the importance of re-examining the TAM.

As ChatGPT use has become a phenomenon of some complexity, some authors have added additional elements, such as ChatGPT-induced anxiety and stress, to synthesize the TAM model.

Based on responses by 23,882 people in 21 countries and in 12 languages

3 Trust in ChatGPT as artificial intelligence technology

The widely spread use of ChatGPT and many other AI tools has brought again the debate about trust in technology and about trust itself. One popular definition explains that trust is “a psychological state comprising the intention to accept vulnerability based upon positive expectations of the intentions or behavior of another”. Since approaches to human trust are different, we cannot cover the whole debate about trust in this paper. However, it is important to note that inconsistency in the terminology have been highlighted. Due to its complexity, trust has been analyzed from many different perspectives; for example, the social dimension of trust; its relationship with well-being; its importance in economics; the economics and biology of trust; and the philosophy and epistemology of trust, just to mention some examples among many other perspectives and abundant bibliography.

It is important to note that trust is one of the elements involved in communication. In fact, the more trust there is in another person, the better the communication can be, to the point that trust is required to create a climate in which honest communication can thrive. Several authors have described different elements as components of trust, such as risk and interdependence, agreed upon across disciplines as conditions that must be present for trust to emerge. In addition, there is a tension between trust and risk, several studies have shown that low risk perception is associated with trust in technology. Taking risks requires trust.

With many technological advancements throughout human history, the concept of trust in technology has also been developed and there are now a number of approaches to building trust in technology. However, some authors are more skeptical on talking about trust in technology. It has been said that “people trust people, not technology”. It is important to note that we cannot talk about the same type of trust between human beings and trust between a human being and a machine or a product of technology. In the case of humans, we speak of integrity, ability, competence, and benevolence; and in the case of technology we speak of reliability, functionality and helpfulness.

While interpersonal trust has been defined as “an expectancy held by an individual or a group that the word, promise, verbal or written statement of another individual or group can be relied upon”; trust in technology has two essential elements: first, technology is considered reliable enough for people to be able to rely on it; and second, that the standard expectation is that the technology is going to perform to a certain level.

In fact, the worldwide use of technology evidences some kind of trust in technological tools, in what has been called ‘intention to use’, otherwise people were not using technology at all or not in the same way it is used globally today. For example, human-like trusting beliefs have been compared to their corresponding trusting beliefs in technology. Janzik and Quandt proposed a model for the definition of trust in media technology as an interaction between several elements: the characteristics of the individual, the social environment, the characteristics of the technology and the specific technology itself. The most recent systematic review of trust in technology has given very useful insights to understand this topic.

It describes the antecedents and consequences of trust in technology, and describe three categories of trust in technology: First, experience with and perceptions of the technological tool; second, the reputation, communication and branding of the company providing the technology; and third, calculative reasoning.

The debate about trust in technology has been enriched by introducing the nuance of trust in AI, as several authors have already done due to the increased development of AI in recent years. Elements such as transparency, performance and explainability have been described as essential for trust in AI but also add that “for the AI system to be trusted by the users, the AI’s trustworthiness must be truly perceived by them”. Moreover, this discussion about trust in AI has been brought to specific disciplines or AI tools. In the case of ChatGPT, a qualitative study analyzing Generation Z’s (born between 1995 and 2012) knowledge and usage of ChatGPT found that trust has been described as accuracy, familiarity, and perceived reliability of content. Some authors have found that trust has an important direct impact on both intention to use and actual use of ChatGPT.

Nowadays, people trust Google and Wikipedia more than ChatGPT, this could be partly explained due to the newness of ChatGPT. However, ChatGPT has capabilities different from those of Google or Wikipedia. A recent literature review summarized the advantages of ChatGPT, such as the ability to generate human-like responses, handle multiple languages, be trained on large datasets, among many others. All these capabilities can lead us to trust in ChatGPT. However, blind trust can lead to overreliance.

4 Overreliance on ChatGPT

Overreliance is defined as “the act or state of relying on something or someone too much: excessive reliance”. A recent study on how an overreliance on AI dialogue systems may affect students’ cognitive skills found that excessive use of AI tools could undermine the practice of critically evaluating information sources, cultivating a deep understanding of research topics, and cross-referencing data, ultimately compromising the ability to perform independent interpretation and analysis. Other authors have warned that overreliance on ChatGPT is one of the concerns about the use of ChatGPT by students in higher education, and Murtiningsih et al. highlight the risk of overreliance on ChatGPT leading to a uniform writing style and potential negative effects on students’ brain development.

In addition, a recent study found that frequency of using ChatGPT predicted reliance on the tool (β = 0.619, p < 0.001), highlighting a habituation effect. A cross-sectional study of 148 students in Morocco that showed that the frequency with which participants used ChatGPT significantly predicted their reliance on it, indicating that longer duration of use corresponded to an increase in reliance, although the study did not fully explain this relationship. ChatGPT could also be described as the “fire of Prometheus” for researchers whose first language is not English when they are involved in scholarly writing. In this review, it is suggested that overreliance on ChatGPT can come from two sources: First, from its excessive use, which could be called overdependency. Second, from disregard or ignorance that there are certain limitations in using ChatGPT.

Overdependency on ChatGPT. Dependency is a potential downside of the use of AI. More specifically, dependency on ChatGPT is a technical dependency that can affect critical thinking skills (Hua et al.. Chakraborty et al., talk about a

ChatGPT Dependency Disorder specifically in healthcare practice, and refer to this disorder as potential overreliance or attachment to ChatGPT. They claim that this disorder raises concerns about the reliance on AI for decision making, possible psychological effects on healthcare professionals and reduced human interaction.

ChatGPT answers are given very quickly, easily replacing human thinking to generate new ideas or thoughts. The fast-paced environments with tight deadlines for submitting work and homework in various types of organizations and companies, including universities and schools, lead people to rely upon quick results over quality results. Second, overdependency on the use of ChatGPT could lead to preference for results of lower quality to those that could be generated in other more reliable and desired higher quality ways can make users get used to low quality answers. Third, there are several personal situations that can lead a person to use ChatGPT and develop overdependency on it. For example, a higher level of academic stress is exhibited by students who have low academic self-efficacy, resulting in augmented expectations of AI technology and augmented levels of AI dependency.

Similarly, a study on the impact of smartphone dependency on real-life leisure activities found that smartphone use behavior, apart from being driven by dependency, is influenced by situational and personal factors.

Disregard for the limitations of ChatGPT. Overreliance is “reliance even in situations where the advice is contrary to available context information and against the advisees’ own interests and those of others”. Cong et al. have warned, for example, that learners may become over-reliant on the AI tool, which could lead to a decline in their ability to think critically. In the context of high-stakes decision making, the consequences of overreliance on, or blind trust in, ChatGPT, cannot be overstated. We propose that overreliance on ChatGPT may also be due to blindness or ignorance of the limitations, disadvantages, and ethical, privacy, and security concerns that lead users to over rely on the AI tool. In Table 2 we summarize limitations, disadvantages, and challenges of ChatGPT from published articles where this AI application is explicitly mentioned.

These are grouped into the following categories: training data, internal functioning, human learning processes, ChatGPT outputs, privacy and security, and sustainability.

Training data is one of the major concerns with ChatGPT given that its training data and the method of obtaining the data has not been fully disclosed. Due to the black box nature of the proprietary program, it is difficult to provide an explanation of how and why questions are answered in a particular way. Other important issues are privacy and security. It is not known what happens with confidential information that is given as a prompt in ChatGPT, for example when a CV is uploaded to be modified for a job application or to create a cover letter for a particular vacancy. Other authors describe limitations in research, scientific literature searches, or academic writing.

Likewise, Memarian and Dolek made a list of 27 issues with using ChatGPT for teaching, including accountability, plagiarism, privacy and the risk of bias and discrimination and found that the most common challenge was misuse or lack of learning mentioned in nine articles, followed by eight articles mentioning the requirement of technical expertise in ChatGPT.

Another disadvantage of ChatGPT that cannot be disregarded is its use of anthropomorphic language, which has raised concerns. This kind of language has several advantages such as giving tailored answers to user’s needs or making dense information more

Table 2 Challenges, limitations, disadvantages, and ethical concerns of the using ChatGPT accessible to users. Moreover, some authors have recently explained the “robust potential of ChatGPT in providing self-help resources and psychoeducation to people with mental health concerns” and described ChatGPT “as a psychotherapist for anxiety disorders”. However, the anthropomorphic language can be seductive, interacting in ways that can be persuasive, manipulative and deceptive, something particularly relevant in topics related to mental health or suicide consultations. Although recent research shows that the suicide information provided by generative AI chatbots has improved in quality and accuracy, this is a sensitive topic that requires ongoing review.

Despite limitations, drawbacks, challenges, and privacy and ethical concerns with ChatGPT, some authors, such as Rice et al. argue that the ultimate responsible parties are the users of ChatGPT, similar to past experiences with calculators and other technological tools. However, this does not mean that ChatGPT users who use it without considering its limitations are overreliant, nor does it relieve ChatGPT developers of responsibility for the negative consequences that such applications may have. Overreliance on ChatGPT can pave the way for deeper problems, such as dependency, where users not only prefer ChatGPT, but begin to rely on it for tasks they could perform independently.

5 AI in health policy: exploring ChatGPT’s potential

AI-based health interventions are being implemented to address various health concerns. In public health, AI is being used for pandemic or epidemic modeling, health diagnosis, disease forecasting, risk prediction, public health surveillance and spatial modeling. In addition, making decisions is potentially thousands of times faster than humans using AI systems. Public health organizations have proposed six priorities about the use of AI in leadership, infrastructure, people, partnerships, equity and AI best practices. In 2023, Ramezani et al. found that policymaking is a very complex process that has several dimensions for which AI can be used, for example at the policy level, at the society level, with policy outputs, among other applications.

Despite the existence of a scoping review about the application of AI in health policy, to our knowledge there is no systematic review on the use of ChatGPT in health policy.

Several studies, from different perspectives, have shown how ChatGPT can be used in health policy. For example, Morita et al. highlight five areas in which ChatGPT can be used: first, policy development and analysis; second, policy communication; third, policy evaluation; fourth, decision-making; and fifth, bias detection and mitigation ChatGPT can also be used in management and planning. And other authors described potential applications, such as policy analysis and development, bias detection, and ethical concerns, among many others.

The capabilities of ChatGPT that have been explored in studies outside the field of health policy could also be applied to this field. For example, ChatGPT can help in conducting literature reviews, developing outlines, and improving writing style in a very short period of time which can facilitate the process of health policymaking. There is also great potential of ChatGPT in disseminating evidence and translating knowledge. ChatGPT is performing even better than some experienced professionals. For example, in clinical medicine, ChatGPT-4 outperformed physicians in diagnosing diseases from medical records, with an average score of 90% compared to physicians with access to ChatGPT (76%) and those without (74%).

The key point of this study was that the AI tool alone outperformed doctors even when they had access to the tool, suggesting that further development of person-computer interactions is needed to realize all the capabilities of AI in decision making processes.

It will be necessary, then, for health policy makers to be trained to use the tool, and analogously as with learning a new language, the only way to learn to use ChatGPT is to be exposed to it and try to use it. Nevertheless, the users of these tools, i.e., health policymakers would be fully responsible of the work made, including omissions or mistakes. Having evidence that AI tools, such as ChatGPT can perform well in health policy, as mentioned before, one of the next steps is promote its acceptance in the public domain where it can be mostly used.

As Dwivedi et al. recently highlighted, it is important to develop tactics to overcome organizational resistance and to promote responsible use of AI among government workers. “Incorporating ChatGPT and other AI technologies in government operations signifies a crucial transition towards a more dynamic, efficient, and responsive public administration (...) AI, such as conversational models like ChatGPT, plays a crucial role in facilitating the change of governments as they progress and adjust to the digital era”. Basic AI literacy should be a priority for those applying it, so they do not take for granted its capabilities and limitations.

Not only to shed light on how AI applications, such as ChatGPT could be used in health policy, but also to promote critical thinking, we propose a set of six requirements for generative AI literacy, based on Annapureddy et al., (See Fig. 1). These elements contribute to build an optimal reliance on ChatGPT avoiding the loss of critical thinking skills and even enhancing them. These could also be asked as mandatory or necessary skills for future jobs of health policymakers using ChatGPT. As a consequence, future candidates for health policy positions might be required to be prepared to use AI tools such as ChatGPT, and government agencies directly involved in health policy would be obliged to provide the necessary training for their employees, as well as training in the ethical and responsible use of this tool.

The most important aspect that will contribute to optimal reliance on ChatGPT will be the development of the necessary skills to interact with the AI, especially the ability to evaluate the results so that better prompts can be designed. Furthermore, the development of checklists will help policymakers to stay on track with the objectives of their policy work, while being aware of the limitations and disadvantages of ChatGPT and avoid privacy, security and ethical issues as much as possible. These checklists could also help to verify the reliability of AI interventions.

6 Conclusions

ChatGPT has been described as a novel conversational AI application with great potential to perform various tasks. Trust in ChatGPT is similar to trust in humans. In this review, we have suggested that as trust in the tool increases, critical thinking skills are likely to increase. However, there is a possibility of overreliance in the tool, and this is when critical thinking may be compromised as the tool is used without consideration of its limitations, drawbacks and ethical challenges.

One potential case for the use of ChatGPT is in health policy, where the tool can have a major impact. However, some basic requirements are suggested for those working in health policy so that they do not lose their critical thinking skills but rather develop critical dialogical thinking. Users have two main opportunities to develop critical thinking skills while they have a dialogic relationship with ChatGPT: first, in the generation of prompts; and second, in the analysis of the responses given by the chatbot. We proposed six requirements that contribute to optimal reliance on ChatGPT to avoid the loss of critical thinking skills and even improve them. We proposed that these requirements could into fundamental skills for some jobs in health policymaking. We suggest that the development of a model of optimal reliance on ChatGPT is urgently needed.

These requirements for future policy makers could also be complemented by structural factors added institutionally to support these new requirements. For example, health policymakers could be required in their work to provide citations of the approach they used and indicate which model of AI, if any, was used to generate the policy. In the end, there may be valuable applications of this, and other technologies provided they are implemented with care.

C.R. and L.S. contributed to the study conception and design. C.R. conducted the material collection and drafted the initial manuscript. C.L., B.I. and L.S. provided new insights, feedback, and revisions to subsequent versions of the manuscript. All authors read and approved the final manuscript.

Funding

No funding was received for this study.

Data availability.

This study is a narrative review and did not generate, collect, or analyse any new datasets. No datasets were obtained from web-based repositories, and therefore no accession numbers, repository names, or direct web links are applicable. The review is based exclusively on previously published peer-reviewed literature and publicly available documents, all of which are cited and accessible through the reference list of the manuscript. As no datasets or supplementary data files were used or produced, no in-text references to data tables or additional files are included.

Declarations

Ethics approval and consent to participate

This study is a narrative review based solely on previously published literature. No human participants or personal data were involved; therefore, ethical approval was not required. Not applicable. This article does not involve human participants.

Consent for publication

Not applicable. The manuscript does not contain any individual person’s data in any form (including images, videos, or case details).

Competing interests

Received: 28 July 2025 / Accepted: 8 February 2026

1. World Health Organization.

Ethics and Governance of Artificial Intelligence for Health Geneva. 2021.

2. European Parliament.

Artificial intelligence act. European Parliament; 2024. ​h​t​t​p​s​:​/​/​w​w​w​.​e​u​r​o​p​a​r​l​.​e​u​r​o​p​a​.​e​u​/​d​o​c​e​o​/​d​o​c​u​ m​e​n​t​/​T​A​-​9​-​2​0​2​4​-​0​1​3​8​​E​N​.​p​d​f.