Background Large language models (LLMs) show promise in clinical contexts but can generate false facts (often referred to as “hallucinations”). One subset of these errors arises from adversarial attacks, in which fabricated details embedded in prompts lead the model to produce or elaborate on the false information. We embedded fabricated content in clinical prompts to elicit adversarial hallucination attacks in multiple large language models. We quantified how often they elaborated on false details and tested whether a specialized mitigation prompt or altered temperature settings reduced errors. Methods We created 300 physician-validated simulated vignettes, each containing one fabricated detail (a laboratory test, a physical or radiological sign, or a medical condition). Each vignette was presented in short and long versions—differing only in word count but identical in medical content. We tested six LLMs under three conditions: default (standard settings), mitigating prompt (designed to reduce hallucinations), and temperature 0 (deterministic output with maximum response certainty), generating 5,400 outputs. If a model elaborated on the fabricated detail, the case was classified as a “hallucination”. Results Hallucination rates range from 50 % to 82 % across models and prompting methods. Prompt-based mitigation lowers the overall hallucination rate (mean across all models) from 66 % to 44 % (p < 0.001). For the best-performing model, GPT-4o, rates decline from 53 % to 23 % (p < 0.001). Temperature adjustments offer no significant improvement. Short vignettes show slightly higher odds of hallucination. Conclusions LLMs are highly susceptible to adversarial hallucination attacks, frequently generating false clinical details that pose risks when used without safeguards. While prompt engineering reduces errors, it does not eliminate them.

Background: Large language models are increasingly evaluated for use in healthcare. However, concerns about their impact on disparities persist. This study reviews current research on demographic biases in large language models to identify prevalent bias types, assess measurement methods, and evaluate mitigation strategies. Methods: We conducted a systematic review, searching publications from January 2018 to July 2024 across five databases. We included peer-reviewed studies evaluating demographic biases in large language models, focusing on gender, race, ethnicity, age, and other factors. Study quality was assessed using the Joanna Briggs Institute Critical Appraisal Tools. Results: Our review included 24 studies. Of these, 22 (91.7%) identified biases. Gender bias was the most prevalent, reported in 15 of 16 studies (93.7%). Racial or ethnic biases were observed in 10 of 11 studies (90.9%). Only two studies found minimal or no bias in certain contexts. Mitigation strategies mainly included prompt engineering, with varying effectiveness. However, these findings are tempered by a potential publication bias, as studies with negative results are less frequently published. Conclusion: Biases are observed in large language models across various medical domains. While bias detection is improving, effective mitigation strategies are still developing. As LLMs increasingly influence critical decisions, addressing these biases and their resultant disparities is essential for ensuring fair artificial intelligence systems. Future research should focus on a wider range of demographic factors, intersectional analyses, and non-Western cultural contexts. Graphic

Background: With their unmatched ability to interpret and engage with human language and context, large language models (LLMs) hint at the potential to bridge AI and human cognitive processes. This review explores the current application of LLMs, such as ChatGPT, in the field of psychiatry.Methods: We followed PRISMA guidelines and searched through PubMed, Embase, Web of Science, and Scopus, up until March 2024.Results: From 771 retrieved articles, we included 16 that directly examine LLMs’ use in psychiatry. LLMs, particularly ChatGPT and GPT-4, showed diverse applications in clinical reasoning, social media, and education within psychiatry. They can assist in diagnosing mental health issues, managing depression, evaluating suicide risk, and supporting education in the field. However, our review also points out their limitations, such as difficulties with complex cases and potential underestimation of suicide risks.Conclusion: Early research in psychiatry reveals LLMs’ versatile applications, from diagnostic support to educational roles. Given the rapid pace of advancement, future investigations are poised to explore the extent to which these models might redefine traditional roles in mental health care.

ChatGPT Health was launched in January 2026 as OpenAI’s consumer health tool and has reached millions of users. Here we conducted a structured stress test of triage recommendations using 60 clinician-authored vignettes across 21 clinical domains under 16 factorial conditions, yielding 960 total responses. Performance followed an inverted U-shaped pattern, with the most dangerous failures concentrated at clinical extremes—nonurgent presentations (35%) and emergency conditions (48%). Among gold-standard emergencies, the system undertriaged 52% of cases, directing patients with diabetic ketoacidosis or impending respiratory failure to 24–48 h evaluation rather than the emergency department, while correctly triaging classical emergencies such as stroke and anaphylaxis. When family or friends minimized symptoms, indicating anchoring bias, triage recommendations shifted significantly in edge cases (odds ratio = 11.7, 95% confidence interval = 3.7–36.6), with the majority of shifts toward less urgent care. Crisis-intervention messages activated unpredictably across suicidal ideation presentations, occurring more frequently when patients described no specific method than when they did. Patient race, sex and barriers to care did not show significant effects, although confidence intervals did not exclude clinically meaningful differences. These findings reveal missed high-risk emergencies and inconsistent activation of crisis safeguards, raising safety concerns that warrant prospective validation before consumer-scale deployment of artificial intelligence triage systems.

Background: The capabilities of large language models (LLMs) to self-assess their own confidence in answering questions within the biomedical realm remain underexplored. Objective: This study evaluates the confidence levels of 12 LLMs across 5 medical specialties to assess LLMs’ ability to accurately judge their own responses. Methods: We used 1965 multiple-choice questions that assessed clinical knowledge in the following areas: internal medicine, obstetrics and gynecology, psychiatry, pediatrics, and general surgery. Models were prompted to provide answers and to also provide their confidence for the correct answers (score: range 0%‐100%). We calculated the correlation between each model’s mean confidence score for correct answers and the overall accuracy of each model across all questions. The confidence scores for correct and incorrect answers were also analyzed to determine the mean difference in confidence, using 2-sample, 2-tailed t tests. Results: The correlation between the mean confidence scores for correct answers and model accuracy was inverse and statistically significant (r=−0.40; P=.001), indicating that worse-performing models exhibited paradoxically higher confidence. For instance, a top-performing model—GPT-4o—had a mean accuracy of 74% (SD 9.4%), with a mean confidence of 63% (SD 8.3%), whereas a low-performing model—Qwen2-7B—showed a mean accuracy of 46% (SD 10.5%) but a mean confidence of 76% (SD 11.7%). The mean difference in confidence between correct and incorrect responses was low for all models, ranging from 0.6% to 5.4%, with GPT-4o having the highest mean difference (5.4%, SD 2.3%; P=.003). Conclusions: Better-performing LLMs show more aligned overall confidence levels. However, even the most accurate models still show minimal variation in confidence between right and wrong answers. This may limit their safe use in clinical settings. Addressing overconfidence could involve refining calibration methods, performing domain-specific fine-tuning, and involving human oversight when decisions carry high risks. Further research is needed to improve these strategies before broader clinical adoption of LLMs.