In a groundbreaking study published in The Lancet Digital Health, researchers from the Icahn School of Medicine at Mount Sinai have illuminated a critical vulnerability in medical artificial intelligence (AI) systems: their propensity to inadvertently propagate falsehoods cloaked in the language of legitimate clinical communication. This revelation underscores an urgent challenge as healthcare increasingly integrates advanced AI technologies intended to enhance the accuracy and safety of patient care through sophisticated data management.
The study meticulously evaluated the responses of nine leading large language models (LLMs) when confronted with medical misinformation embedded in realistic texts. These texts included hospital discharge summaries, social media posts from platforms such as Reddit, and meticulously crafted clinical vignettes verified by medical professionals. The researchers engineered each scenario to contain a single fabricated medical recommendation, deliberately camouflaged within authentic clinical or patient communication styles to test the resilience of these AI systems against disinformation masked as factual guidance.
One striking example within the study exposed the dangerous consequence of this susceptibility: a falsified medical discharge note advised patients suffering from esophagitis-related bleeding to “drink cold milk to soothe symptoms.” Rather than flagging this spurious advice as unsafe or inaccurate, multiple LLMs accepted it unquestioningly, treating the fabricated statement with the deference typically reserved for validated clinical recommendations. This acceptance highlights a systemic flaw where the AI’s trust in language patterns supersedes the factual correctness of the content.
According to Dr. Eyal Klang, co-senior author and Chief of Generative AI in the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai, the findings reveal a worrying trend. These AI systems default to interpreting confident and familiar clinical language as truth, irrespective of the underlying veracity. In essence, the models prioritize linguistic presentation over factual integrity, which could enable the silent circulation of medical misinformation through digital healthcare channels.
The crux of the problem lies in the models’ training processes. LLMs learn from extensive datasets that often amalgamate vast quantities of textual data without an intrinsic mechanism for validating factual content. Consequently, when false information mimics the stylistic features of authentic medical documents or patient discussions, the models lack the critical tools needed to discern and challenge inaccuracies effectively.
To rigorously quantify this vulnerability, the research team devised a large-scale stress-testing framework. This paradigm systematically measured the frequency and contexts in which AI models ingested and regurgitated false medical claims, whether presented neutrally or embedded within emotionally charged or leading phrasings typically used in social media environments. These nuanced linguistic variations influenced the AI’s propensity to accept or reject misinformation, indicating that even subtle changes in expression can sway model responses.
Given these insights, the authors advocate for a paradigm shift in how AI safety in clinical settings is approached. Rather than assuming AI systems are inherently reliable, they emphasize the imperative to develop measurable metrics that assess an AI’s likelihood to “pass on a lie” before deployment. Integrating such metrics into AI validation pipelines could serve as a crucial checkpoint in protecting patient safety and preserving the integrity of medical information.
Dr. Mahmud Omar, the study’s first author, underscores the practical implications of this approach. By utilizing the dataset created through their research as a benchmarking tool, developers and healthcare institutions could systematically evaluate the robustness of existing and next-generation medical AI models. This proactive evaluation strategy could substantially reduce the risk of false medical advice disseminated through automated systems.
The collaborative efforts leading this research involve a multidisciplinary team spanning clinical medicine, data science, and digital health innovation, suggesting a comprehensive approach to the ethical use of AI in healthcare. Their work aligns with the broader mission of the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai, which pioneers responsible integration of AI in medicine—ensuring these technologies augment rather than undermine clinical decision-making.
The ramifications of this study extend beyond simply identifying faults; they ignite a call for instituting built-in safeguards within AI-powered clinical support tools. Mechanisms such as real-time evidence verification, contextual uncertainty estimation, and cross-referencing with trusted medical databases may form the foundation of future AI architectures that proactively filter out misinformation and alert clinicians to questionable inputs.
Furthermore, these findings raise compelling considerations about the interplay between AI and the ever-evolving landscape of digital health communication. As patient care increasingly incorporates inputs from social media and other informal sources, AI systems stand at the convergence of potentially conflicting data streams. Ensuring their ability to reliably discern credible information is paramount to preventing inadvertent harm.
Looking ahead, this research sets a new benchmark for evaluating AI tools in healthcare, challenging the community to prioritize not just functionality but veracity and safety. The framework established by the researchers will likely be instrumental in guiding regulatory standards, industry best practices, and future academic inquiry into the responsible deployment of AI in medicine.
As AI technologies become more pervasive in clinical workflows, from diagnostic aids to patient education, the integrity of their outputs must be beyond reproach. This study’s spotlight on the susceptibility of language models to medical misinformation underscores a vital frontier where AI ingenuity must be coupled with rigorous safeguards to truly transform patient care outcomes beneficially.
Subject of Research: People
Article Title: Mapping LLM Susceptibility to Medical Misinformation Across Clinical Notes and Social Media
News Publication Date: 9-Feb-2026
Web References: https://icahn.mssm.edu/about/artificial-intelligence
References: The Lancet Digital Health, DOI: 10.1016/j.landig.2025.100949
Keywords: Generative AI, Medical misinformation, Large language models, Clinical AI, Healthcare technology, AI safety

