Study Finds LLM Confidence Tone Does Not Correlate with Correctness

ClickOk5811 · reddit · 2026-08-28

An analysis of 40 classification outputs compared hedged phrasing ("likely," "probably") with flat, confident statements against ground truth. Results showed wrong answers occurred about one in six times for both styles. Hedge words did not track actual uncertainty but were more stylistic. Relying on a model's confident tone as a trust signal is unreliable.

Original post →

More from Models

Models channel →