Quantifying and mitigating premature closure in frontier LLMs.

Premature closure—reaching a conclusion before enough information is available—is a known source of diagnostic error and may also affect LLM output. This preprint study tested five advanced LLMs on structured and open-ended medical questions, including challenging physician-designed cases. The models often answered despite unresolved uncertainty. Although safety-oriented prompts reduced these errors, important failures remained, highlighting the need to assess not only what medical AI systems know, but also whether they can recognize when to withhold judgment.