A recent study published in BMJ Open reports on Apr. 16 that nearly half of responses from widely used free artificial intelligence (AI) chatbots to health-related questions were problematic or failed to align with scientific consensus. The audit evaluated five public-facing AI models using 250 prompts across categories prone to misinformation, including cancer, vaccines, stem cells, nutrition, and athletic performance.
The study matters because as AI chatbots become increasingly common in daily life—used by an estimated 75% of workers for routine tasks—their role in providing health information is under scrutiny. Many users rely on these tools for medical advice, raising concerns about the safety and accuracy of their responses.
Researchers found that while overall performance did not differ significantly among Gemini 2.0, DeepSeek V3, Llama 3.3, ChatGPT 3.5, and Grok 2 models (p = 0.566), a total of 49.6% of chatbot answers were classified as somewhat or highly problematic by subject-matter experts evaluating the outputs against predefined criteria. The authors noted distinct behavioral vulnerabilities among individual models such as poor citation authenticity and difficult readability levels equivalent to college sophomore-to-senior grades.
Open-ended questions were more likely to generate highly problematic responses (32%) compared to closed-ended ones (7.2%). While vaccine- and cancer-related prompts produced fewer issues than expected by chance alone, nutrition and athletic performance topics resulted in higher rates of inaccurate information.
Citation quality was low across all platforms; only about 40% of requested references were provided completely or accurately on average per model—with Gemini returning the fewest citations overall—and readability scores indicated content was challenging for most readers.
The study concludes that users should be extremely critical when seeking medical advice from AI chatbots and should consult human specialists before acting on any recommendations provided by these technologies. The authors also recommend increased public education and oversight regarding the use of health-related AI tools.
