Judge-dependent safety gains and model-specific helpfulness costs of evidence-sufficiency prompting in clinical LLMs

En 30 secondes

Ce qui s'est passé

A prompt technique reduced unsafe overconfidence in clinical AI models from 49% to 25%, but the safety gain heavily depends on which AI evaluates the results.

Pourquoi ça compte

Engineers building clinical AI systems need to know that safety metrics from AI judges may not reflect real behavior change and require human validation.

Vigilance

The measured safety improvement vanished by half when a different AI judge scored the same answers, and helpfulness dropped severely for some models but not others.