ChatGPT Health delivers inaccurate health assessments from Apple Health data
ChatGPT Health reportedly delivered inaccurate and alarmist health assessments when analyzing ten years of Apple Watch data from a test user, including false cardiovascular risk warnings that a physician contradicted, and gave inconsistent results across multiple consultations with the same data.
What happened
ChatGPT Health, OpenAI's feature that connects Apple's Health app with the chatbot, was tested by The Washington Post analyst Geoffrey A. Fowler using ten years of accumulated health data including 29 million steps and 6 million heart rate measurements.
The system assigned a failing grade ('F') to Fowler's health despite no actual medical risks, later revising to 'D' when additional information was provided. The AI generated inconsistent conclusions, including false suggestions of cardiovascular risk, and forgot basic user data such as age and gender when consulted multiple times.
Reported impact
A physician reviewing Fowler's data confirmed he had excellent health and required no additional tests or treatment, contradicting ChatGPT Health's assessments.
The inconsistency and alarmist conclusions risk misleading users into making health decisions based on unreliable AI recommendations.
Index analysis
ChatGPT Health relies on estimated metrics like VO₂ max that smartwatches do not measure with clinical precision, contributing to inaccurate analysis.
The AI system failed to maintain consistency in its analysis, forgetting fundamental user information between separate consultations with the same data.
Governance controls
OpenAI stated that ChatGPT Health is not designed to diagnose or treat diseases. However, the tool is not subject to strict medical regulations such as HIPAA, despite handling sensitive health data.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.