Five Widely Used AI Chatbot Generate Misleading Health Answers

(MENAFN) A new study has found that five widely used AI chatbots frequently generated misleading or problematic responses when asked medical and health-related questions.

Researchers publishing in BMJ Open on Tuesday evaluated Gemini, DeepSeek, Meta AI, ChatGPT, and Grok using 50 prompts designed to trigger misinformation across high-risk health categories.

The topics included cancer, vaccines, stem cells, nutrition, and athletic performance, with questions deliberately structured in February 2025 to push the systems toward unreliable or misleading guidance.

Out of 250 total responses analyzed, 49.6% were judged problematic. This included 30% classified as somewhat problematic and 19.6% rated highly problematic.

While overall performance differences between the chatbots were not statistically significant, Grok stood out for producing a higher proportion of highly problematic answers compared with the others.

According to the findings, the strongest performance appeared in vaccine and cancer-related queries, while the weakest results were seen in stem cell, nutrition, and athletic performance topics. Researchers also noted that open-ended questions led to significantly worse outputs than closed-ended ones.

Citation quality was another major concern. Across 25 closed-ended prompts per chatbot, the systems returned roughly 81% of requested references, but the median completeness score dropped to just 40%. None of the chatbots produced a fully accurate and complete reference list.

The study further found that all five systems generated responses that were difficult for the average user to understand, with readability levels generally suitable only for individuals with higher education.

Researchers warned that continued use of AI chatbots in healthcare contexts without stronger oversight could accelerate the spread of medical misinformation.

MENAFN20042026000045017169ID1111005753