Study: AI Chatbots Misdiagnose Medical Cases in Over 80% of Scenarios With Incomplete Information
A study by Mass General Brigham tested 21 language models from major AI companies and found that chatbots like ChatGPT and Gemini misdiagnosed medical cases in over 80% of scenarios when given incomplete patient information, and had error rates exceeding 40% even with complete data due to AI hallucinations.
What happened
Mass General Brigham from Massachusetts tested 21 language models developed by OpenAI, Anthropic, Google, xAI, and DeepSeek on 29 clinical scenarios constructed from medical reference texts.
Chatbots were given stepwise information about patients including symptoms, consultation observations, and laboratory test results. All tested models provided incorrect diagnoses in over 80% of cases when information was incomplete, and had error rates exceeding 40% even with complete patient data.
Reported impact
The study found that chatbots including ChatGPT and Gemini can lead users down a dangerous path by providing incorrect medical diagnoses. Experts concluded that while chatbots can be useful for general information, they cannot replace medical evaluation by a physician.
Index analysis
Experts explained that the accuracy of responses depends largely on the volume and quality of information received. All current language models suffer from AI hallucinations, where they generate responses that seem plausible but are not supported by real or sufficient data.
Performance varied: in some situations certain models correctly identified diagnoses in up to 90% of cases, but consistency was not maintained across all scenarios.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.