Artificial Intelligence and Health Literacy: Possibilities and Limitations of Publicly Accessible AI Language Models

The use of artificial intelligence (AI) has become indispensable in healthcare. AI is already widely used in the field of health education, e.g.
Chatbots integrated into health apps are being used (Aggarwal et al. 2023) to promote health literacy. A high level of health literacy, defined as the ability to access, understand, and apply health information (WHO 2019), is crucial for making informed health decisions (Nutbeam 2008). This competence also includes the ability to distinguish credible health information from misinformation in the digital age (Debad/Holmes 2024). This is particularly important as traditional health-related information sources (e.g., medical professionals, TV, print and online media) are being supplemented by AI-based information services such as ChatGPT. ChatGPT was first made publicly available by OpenAI in November 2022 and quickly gained popularity, especially among younger age groups (Kasneci 2024). The extent to which such tools influence citizens' health literacy remains to be seen. It is also unclear to what extent access to health information is facilitated compared to classic internet research („Dr. Google”) and what new challenges arise regarding the quality and interpretation of this information.

High health literacy empowers individuals to participate in decision-making processes with healthcare providers, thereby gaining greater control over their own health (Caeiros et al. 2024). High health literacy is associated with better health outcomes and reduced inequality, as individuals acquire the knowledge to manage chronic conditions and navigate the healthcare system (Caeiros et al. 2024). It can serve as a resource against adverse social conditions and promote more equitable access to healthcare (Zanobini et al. 2024). Despite the importance of health literacy, many people face barriers that make it difficult for them to achieve sufficient health literacy. These include limited access to reliable information and insufficient media literacy skills to cope with the flood of digital information. AI-based language models (e.g., ChatGPT) enable users to quickly access extensive health information and thus understand complex issues. However, they can also lead to inaccurate or incorrect answers (so-called hallucinations) (Vij et al. 2024). Furthermore, there is initial evidence that while ChatGPT provides users with different demographic profiles with the same information, it is formulated differently (Andreadis et al. 2024), which could lead to different interpretations of the AI responses.
Regarding public perception of AI in healthcare, a study found that over 901,000 respondents had heard or read about AI, but only 241,000 had good or very good knowledge in this area. Overall, around half (53,181,000) of the respondents rated the use of AI in medicine as positive or very positive, while only 4,771,000 held a negative or very negative view. Despite a lack of major concerns, the majority believed that the new technology should only be used under medical supervision. Older patients, women, and individuals with lower levels of education and low levels of technical proficiency were particularly cautious (Fritsch et al. 2022).
It is striking that previous research has been largely limited to AI-based education for healthcare professionals and academic education. Research findings on health literacy in the general population via AI are still scarce. Therefore, this article uses ChatGPT as an example to examine the possibilities and limitations of freely accessible AI-based language models (so-called Large Language Models, or LLMs) for promoting health literacy in the general population.
Three objectives are pursued: First, the potential and limitations of AI systems in providing health-related information are to be identified. Second, the trustworthiness and influence of AI-supported information on users' health literacy are critically examined. Third, based on the findings, recommendations for strengthening digital health literacy through the responsible use of AI technologies are to be derived.
methodology
To investigate the research topic, a narrative literature review was conducted in September 2024 to gain an overview of the subject. For reasons of topicality, the analysis period was limited to the years 2022 to 2024. Traditional literature databases (PubMed) as well as AI-based literature review programs (Consensus, ResearchRabbit, typeset_io) were used for the analysis. The search terms were "artificial intelligence" and "health literacy".
Of 136 identified sources, 25 sources were ultimately included in the analysis after screening (abstracts, full text) (Fig. 1).
Furthermore, the use of traditional literature databases as well as AI-based literature analysis tools should be considered to determine the added value that combining both approaches can bring to scientific research. It turns out that while the majority of exclusions originate from AI-supported literature searches, the overall yield of these searches was significantly increased (14 typeset_io, 9 Consensus, 2 PubMed). This finding aligns with current studies on AI-generated literature searches, which also emphasize that, on the one hand, AI-supported automation can increase the effectiveness of literature search processes, while on the other hand, critical review of the data by researchers is essential. AI-generated literature analyses are particularly well-suited for large datasets and interdisciplinary research topics (Ejjami 2024, Tomczyk et al. 2024). As Figure 2 reveals, publications focusing on quality assurance (content validation, fact verification, and consistency checks) of AI-generated content represent by far the majority of the identified sources.
The following evaluation criteria were used as the basis for the analysis, which were specified on the basis of theoretical considerations and discussion among the authors:
1. Acceptance: Acceptance refers to trust and willingness to use ChatGPT, taking into account target group-specific differences (age, class, etc.) as these determine whether and how (critically) AI-based tools are used.
2. Quality: The quality and comprehensibility of AI-provided information are relevant for (perceived) reliability, acceptance, and the development of health literacy.
3. Barriers: Barriers, including data protection aspects, relate to access to and trust in AI-based tools and in turn influence acceptance and effectiveness.
Results
1. Acceptance and trust
Only a few studies (7 out of 25) addressed the topic of acceptance and trust. Overall, the public tends to trust physicians more than AI technology and is more likely to follow their recommendations than those of an AI, even if these recommendations contain cultural biases. Regarding the question of whether a physician should use AI, public opinion is almost equally divided between support and opposition. There is unease about AI being informed about health status, both now and in the future (Rojahn et al. 2023). Furthermore, particularly when users feel overwhelmed, unease is reported due to the complexity and lack of control (Chalutz-Ben Gal 2023). On the other hand, AI responses are sometimes rated as more empathetic (Ayers et al. 2023b), although these could exacerbate existing problems.
In particular, the abundance of deficient information and AI outputs could unsettle patients with higher socioeconomic and cultural resources, especially since these individuals are already more hesitant to interact with healthcare professionals. For patients with higher resources, easy access to health information can increase anxiety about making the "right" decision (Lautrup et al. 2023). Feelings of respect, a positive attitude, perceived benefits, ease of use, recommendations from others, and existing resources (media literacy, etc.) influence usage (and frequency) (Li et al. 2023; Budler et al. 2023). Women are less convinced of the health benefits of ChatGPT than men. Individuals with higher levels of education and those who have already heard of ChatGPT also tend to rate its health benefits rather low. Health literacy and technological affinity do not have a significant impact on perceived benefits (Platt et al. 2024). However, tech-savvy and open-minded individuals are more likely to use AI-based applications frequently (Ben-Gal 2023).

2. Quality and comprehensibility
Of the 25 publications included, 19 addressed the quality or comprehensibility of information generated by artificial intelligence. Two of these were literature reviews, while the remaining 17 contained original studies. All studies focused on LLMs, primarily ChatGPT in its available English-language versions. The studies either simulated a question-and-answer format or challenged the LLM to provide explanatory information or improve its comprehensibility. Regarding the accuracy of the information, most studies concluded that the LLMs largely provided correct information, but that this information was often generic and contained gaps (Budler et al. 2023; Grünebaum et al. 2023; Hillmann et al. 2024; Karakas et al. 2023; Kassab et al. 2024; Sciberras et al. 2024; Walker et al. 2023). The information provided by LLMs was nevertheless sometimes more accurate than that provided by physicians (Ayers et al. 2023a) and more relevant, albeit more incomplete, than internet searches (Pascual-Presa et al. 2024). More advanced LLMs (e.g., ChatGPT 4.0 compared to ChatGPT 3.5) showed higher accuracy compared to previous versions (Wang et al. 2023; Washif et al. 2023). The quality of the prompting sometimes affected the quality of the responses (Deiana et al. 2023; Lautrup et al. 2023). Furthermore, so-called "actionable" information, i.e., recommendations regarding behavior or access to specific services, was rare (Anaya et al. 2024; Ayers et al. 2023b). Comprehensibility was sometimes described as mixed or limited (Ghanem et al. 2024; Mondal et al. 2023), unless it was explicitly stated that the information should be easily understandable (Abreu et al. 2024; Rouhi et al. 2024). Furthermore, LLM responses are probabilistic, meaning they are based on probabilities to generate the most relevant answer. Therefore, response quality varies from case to case (Ayers et al. 2023).

3. Barriers, data protection and data security
Data protection and data security are addressed in only a few studies. Barriers can arise from the disadvantages faced by certain subpopulations. According to Karakas et al. (2023), while ChatGPT does not store or retrieve personal data, personal health data should not be entered during interaction. Concerns regarding data protection and the willingness to share personal information play a role in the willingness to use AI (Chalutz-Ben Gal 2023; Platt et al. 2024). A facilitating factor for AI use can be the feeling of privacy and respect (Li et al. 2023). In a study by Platt et al. (2024), belonging to an ethnic minority was associated with positive expectations of ChatGPT as a healthcare resource, and in the study by Rojahn et al. (2024), respondents believed that AI might make less culturally biased decisions than human interaction partners. Few studies examine the role of gender as an access factor. Washif et al. (2023) observed that strength training programs generated via ChatGPT for women were more tailored to men. As described in point 1, female gender and a higher level of education are associated with greater reservations about using AI as a health resource and can thus also represent a barrier to its use. An anticipated high effort required for use and resentment towards AI also proved to be barriers to the use of chatbots for health-related questions (Li et al. 2023).
Discussion and limitations
Acceptance and trust appear to depend heavily on personal characteristics (e.g., age, gender, affinity for technology), ethical factors (e.g., privacy, data protection), and technological factors (e.g., usability, type and quality of output). The quality of the output is highly dependent on user input and often appears accurate, but also incomplete. It should be noted that quality assessments often test readability in English, which calls into question the transferability of the results to a German-speaking context. Findings regarding comprehensibility vary considerably. How different user groups interpret and potentially implement AI outputs remains unclear. Ethical aspects (e.g., data protection, data security, privacy) are currently given little consideration (specifically for target groups). The tendency, observed in several studies, for AI to formulate health information at a high level of education can also contribute to exacerbating social inequalities regarding health opportunities and should be considered an ethical challenge. Furthermore, another ethical question arises: If more advanced large language models, which are often fee-based, provide better information, access to this improved content is limited or only available to those with paid subscriptions. Conversely, AI could improve access to health information for minorities due to its low-threshold availability and perceived cultural neutrality. Regarding methodology, a key limitation is the small number of qualitative and quantitative studies with relatively small sample sizes. This is because AI is still a relatively new tool, and quantitative studies in particular take a considerable amount of time (Castillo-Martínez et al. 2024). The comparability and interpretation of studies are further complicated by the significant differences in study design, duration, population groups, and subject matter. Additionally, there is a lack of longitudinal studies (Platt et al. 2024), which, given the rapid pace of AI development, poses a major challenge and can lead to biases. Given the complexity of LLMs, more qualitative and quantitative studies are needed to better understand people's underlying attitudes.
Regarding AI as a literature analysis tool, it should be noted that the results are only partially reproducible, as AI technology is constantly evolving and using different AI tools requires a certain level of understanding. The Effi-
While the efficiency, accuracy, and inclusivity of the literature search process are increased (Ejjami 2024), the high number of irrelevant results demonstrates the need to verify AI results (Fig. 1). AI can be a valuable tool for source acquisition, but it can also lead to inaccuracies in the interpretation of data (AI summaries), making extremely careful source verification essential.
Conclusion
In summary, AI systems can help the public access health information in an understandable way. This can contribute to promoting health awareness, reducing the risk and spread of infectious diseases (e.g., COVID-19), alleviating psychological stress and anxiety, and thus fostering an optimistic outlook (Wang et al. 2023). However, it is important to consider that this could exacerbate health inequalities, as digital health literacy (prompting skills, critical handling of digital information) is less developed, particularly among those with lower levels of education. Similarly, with health apps, a healthy user effect is to be expected, meaning that target groups with a high level of health literacy are more likely to benefit from AI. Overall, the findings also point to a need for action in medical care (keyword: empathy, discrimination) and indicate that greater public awareness and education of healthcare professionals are necessary, as AI-based research (on the part of both the public and researchers) requires a thorough review process and corresponding expertise to ensure the responsible use of AI tools. Future studies should give greater consideration to target group-specific and ethical aspects and be designed as longitudinal studies to better capture long-term effects across different population groups.