Large language models are increasingly shaping how patients understand symptoms, select clinicians and make decisions about investigations and treatments. A recent commentary published in Eye warns that this growing influence creates a new pathway for advertising and commercial manipulation within healthcare.
Unlike traditional search engines, which present users with multiple ranked sources, large language models often synthesise information into a single, confident response. While convenient, this format can make it difficult for patients to distinguish independent clinical evidence from information influenced by marketing, commercial prominence or biased online content.
The authors highlight the risk of “LLM poisoning”—the deliberate flooding of the online information environment with promotional articles, reviews, exaggerated credentials and AI-generated content. If produced at sufficient scale, this material could influence how an AI system recommends clinicians, hospitals, investigations or treatments, even when the underlying claims are not supported by clinical merit.
This issue is particularly relevant in specialties such as ophthalmology, where patients commonly self-refer, seek second opinions and make time-sensitive decisions. An apparently neutral answer to questions such as “Who is the best retinal surgeon?” or “Where should I have cataract surgery?” may reflect online visibility and engineered prominence rather than objective quality.
The concern extends beyond overt advertisements. Commercially framed information may influence patients to request branded therapies, favour experimental treatments or distrust established evidence-based care. As AI-generated answers become more fluent and authoritative, these underlying influences may become even harder to recognise.
The authors call for clearer source attribution, transparent labelling of sponsored content, stronger fact-checking systems and regulatory oversight. For healthcare professionals, the key message is that AI-generated medical recommendations should not be assumed to be commercially neutral.
As conversational AI becomes embedded in healthcare, protecting the information environment that shapes its answers will be as important as assessing the accuracy of the models themselves.
Read more: https://www.nature.com/articles/s41433-026-04518-w