The Knowing-Saying Gap: When Probes See Errors that Confidence Misses
The paper 'The Knowing-Saying Gap' examines a fundamental disconnect in AI models where internal mechanisms detect errors but fail to reflect them in output confidence. Researchers found that internal probes identify corruption with high accuracy, yet the model's verbalized certainty remains constant. This gap is observed across multiple model families and challenges assumptions about how to monitor and improve AI reliability.