More than 40 million people ask health‑related questions to ChatGPT each day, e.g., “Does laptop posture cause a stiff neck and persistent headache?” While posture is plausible, meningitis is a time‑critical alternative that must not be missed. To test whether models exhibit the proper response behavior—epistemic independence—in such scenarios, we introduce the SelfDx benchmark. It consists of 100 paired cases: a caution case where a serious condition remains possible and the model should stay cautious, and an agreement case where the user supplies evidence that rules out the serious alternative, making it safe for the model to agree.
Across 27 frontier models, we observe a systematic tendency to inappropriately agree with the user’s self‑diagnosis in caution cases. Even the best current‑generation model achieves only a 77.0% appropriate caution rate on time‑critical conditions. We experiment with three prompt variants that raise caution but reduce appropriate agreement by 9‑35%, indicating excessive caution. In contrast, supervised fine‑tuning on a balanced set of 128 examples improves both caution and agreement rates simultaneously across three runs.
We recommend post‑training fine‑tuning to boost appropriate caution without making the model overly cautious in every instance, and we advocate paired evaluation to verify both response behaviors.
Review