Chatbots give risky cancer advice

Popular AI chatbots sometimes suggested unsafe cancer care instead of sticking to standard treatment guidance. (nbcnews.com) Researchers published the audit Tuesday in *BMJ Open* after testing ChatGPT, Google Gemini, Meta AI, Grok and DeepSeek with 250 prompts across cancer, vaccines, stem cells, nutrition and athletic performance. They ran the tests in February 2025 and had two subject-matter experts in each category rate every answer. (bmjopen.bmj.com) Nearly half of all answers, 49.6%, were rated problematic: 30% somewhat problematic and 19.6% highly problematic. The paper said Grok produced significantly more highly problematic responses than expected, while overall differences among the five bots were not statistically significant. (bmjopen.bmj.com) The researchers designed the prompts to mimic how people actually search when they already suspect a false claim is true. NBC News reported examples included whether 5G or antiperspirants cause cancer, which vaccines are dangerous, and where to find alternatives to chemotherapy. (nbcnews.com) Large language models work by predicting plausible next words, not by checking facts the way a doctor checks a chart or a guideline. An Oxford-led study published in February found people using these systems for medical scenarios did not make better decisions than people using standard searches or their own judgment. (ox.ac.uk) The BMJ Open paper found the bots were especially hard to trust because they answered with confidence and almost never refused. Out of 250 questions, the chatbots declined to answer only two times. (bmjopen.bmj.com) The study also found the systems did relatively better on vaccines and cancer than on stem cells, athletic performance and nutrition. Even in the stronger categories, the authors said answers were often incomplete, weakly sourced or open to subjective interpretation. (bmjopen.bmj.com) That lands as tech companies push deeper into health tools. OpenAI said in January that ChatGPT Health can connect to medical records and wellness apps, but said the product is meant to support care and “is not intended for diagnosis or treatment.” (openai.com) Google posts a similar warning for Gemini Apps, telling users not to rely on its responses as medical or other professional advice and to double-check outputs because they may be inaccurate or inappropriate. (support.google.com) The new audit does not say every chatbot answer is wrong. It says a confident answer in a high-stakes setting can still be incomplete enough to send someone with cancer or another serious condition in the wrong direction. (bmjopen.bmj.com)