#uncertainty-quantification

#uncertainty-quantification

Evaluation incentives push AI to guess confidently instead of expressing uncertainty, improving test performance but increasing harmful hallucinations and undermining safety and user trust.