As artificial intelligence becomes increasingly woven into everyday life, its presence in healthcare is expanding at an even faster pace. Yet public trust, expectations, and concerns surrounding AI remain unevenly understood. A new Open Access narrative review in The European Journal of Public Health addresses this gap by analyzing how people perceive the use of large language models when seeking health information. The authors examined 120 studies published through May 2024, capturing the perspectives of patients, citizens, healthcare workers, and students across multiple medical fields and geographic regions.
The rapid growth of ChatGPT and similar tools has created a new dynamic in which individuals routinely ask AI questions about symptoms, treatments, test results, or new medical developments. According to the review, many citizens describe these systems as convenient, fast, and easier to understand than traditional medical language. For individuals who struggle to interpret clinical terminology or who have limited access to healthcare professionals, AI tools can reduce barriers to essential information. This expanded access is viewed as particularly valuable in general medicine, where many studies reported that AI responses improved clarity, organization, and emotional tone when explaining medical concepts.
At the same time, the review found a clear recognition that AI systems cannot replace clinicians. Benefits reported by participants often focused on empowerment: a better understanding of one's health, more confidence in navigating medical conversations, and an increased feeling of control. AI tools were also seen as potentially useful for screening tasks, suggesting possible diagnoses, or enhancing comprehension of imaging and lab reports. These advantages, however, were consistently accompanied by acknowledgment that final decisions and interpretations should remain under medical supervision.
Concerns emerged as a central theme across the 120 included studies. Many respondents worried about misinformation, especially in contexts where AI-generated answers appear confident yet contain errors or outdated recommendations. Several included studies documented examples of hallucinated responses - answers that sounded plausible but were incorrect in ways that could mislead patients. Others described cases in which AI-generated guidance did not align with clinical standards, emphasizing that strong performance in everyday conversation does not guarantee medical accuracy.
Privacy also surfaced as a major issue. The review highlights widespread apprehension about how LLMs might store, transmit, or inadvertently expose personal medical data. Participants expressed uncertainty about the security of cloud-based systems, noting that breaches or misuse of sensitive information could undermine trust. Concerns extended to inferred data: even when users do not directly provide medical information, the way they phrase symptoms or concerns could reveal identifiable health patterns.
Another prominent concern involved bias. Because LLMs learn from large quantities of text, they can reflect and reproduce the inequities present in their training data. Several studies referenced in the review found evidence of racial, cultural, or gender-related bias in AI outputs. Citizens and patients who were aware of these issues worried that biased information could reinforce existing disparities in care or produce misleading guidance for marginalized groups.
These risks led many studies in the review to emphasize the need for standardized safeguards. Participants consistently argued that AI systems used for healthcare information should be trained on medically validated, high-quality datasets. They also supported the involvement of clinicians throughout the design, testing, and implementation processes to ensure that AI complements medical expertise rather than undermining it. Transparency was another recurring theme: users wanted to understand where information originated, how reliable it was, and whether clinical experts had reviewed it.
Despite these concerns, the review underscores a generally positive outlook toward AI when used responsibly. Many participants described AI as helpful for preparing questions before appointments, interpreting basic results, or understanding treatment options. Physicians themselves, according to several included studies, see potential for AI to reduce administrative burdens and streamline communication, freeing time for more meaningful patient interaction. However, they also stressed that AI should support - not replace - the human relational aspects of care.
The patient - provider relationship was a key focus of the review. Several studies noted that while AI can generate polite or empathic language, users can distinguish between "artificial empathy" and genuine human understanding. Respondents emphasized that trust, reassurance, and nuanced emotional support remain central to clinical encounters. Many expressed concern that excessive reliance on AI could weaken these fundamental elements of care or shift responsibilities away from trained professionals.
The review also addresses the risk of professional deskilling. If clinicians begin relying on AI systems to guide complex decisions or to interpret nuanced patient messages, there is a concern that essential skills - critical thinking, diagnostic reasoning, and communication - may erode over time. This possibility intensifies the need for careful integration strategies that maintain physicians' active role in health decision-making.
The authors note that the global distribution of studies in the review may produce some geographic bias, as most research originates from the United States and Europe. Nevertheless, the included studies reveal consistent themes across regions. The rapid evolution of LLM technologies also poses a challenge, as findings may become outdated quickly. Even so, the review provides a detailed snapshot of public sentiment at a pivotal moment in AI adoption.
Seen from the perspective of Seven Reflections' Dimensional Systems Architecture, the public's responses reflect a system attempting to reorganize itself around new informational structures. LLMs introduce a parallel cognitive layer - an externalized field of compressed health knowledge. When this field is stable and validated, it enhances clarity and reduces cognitive load. When it is unstable or opaque, it increases uncertainty and disrupts trust. The tension identified in the review mirrors a structural negotiation between human judgment and algorithmic guidance: an evolving boundary that must remain flexible, transparent, and ethically anchored to preserve coherence within the healthcare ecosystem.
The review concludes that large language models have significant potential to improve health information access, reduce disparities, and enhance communication. Yet the risks are equally clear. Without appropriate safeguards - clinical supervision, rigorous data standards, transparency, and privacy protections - AI could compromise patient safety or weaken essential human relationships in care. The authors argue that LLMs should be integrated as complementary tools, enriching but not replacing human expertise. The future of AI in healthcare, according to the review, will depend on maintaining this balance.
