Accuracy and Limitations of ChatGPT and Google Gemini for Health Information: A Mini-Review

Abstract Introduction: Large language models (LLMs) such as ChatGPT and Google Bard—later updated to Google Gemini—are increasingly used for health information, raising concerns about accuracy and clinical safety. This mini-review synthesizes recents evidence on their performance across medical domains. Methods: A rapid-review search identified ten peer-reviewed empirical studies published between 2023 and 2025 evaluating ChatGPT, Google Bard, or Google Gemini on health- and clinical-information tasks. Data on clinical domain, accuracy metrics, performance scores, and reported limitations were extracted. Owing to methodological heterogeneity, findings were synthesized narratively. Results: Accuracy varied substantially by task and domain. LLMs performed best on factual and patient-education queries, especially in dermatology and general disease explanations. Performance declined for multi-step reasoning, drug-therapy decisions, guideline-based recommendations, and differential diagnosis. Across comparative studies, ChatGPT—particularly GPT-4—generally outperformed Bard. All models produced incomplete or inconsistent responses, with documented instances of hallucinations, fabricated citations, outdated or non–guideline-concordant recommendations, and variability across repeated runs. Conclusions: LLMs can provide accessible general health information but remain unreliable for tasks requiring clinical reasoning or up-to-date guideline adherence. They should be used only as adjunct informational tools under human supervision. Continued model refinement, standardized evaluation methods, and safety-focused oversight are essential for responsible integration of AI-generated medical content into health communication.