Epistemic Observability in Language Models

Tony Mason, Vaastav Anand

We find that models report highest confidence precisely when they are fabricating. Across four model families (OLMo-3, Llama-3.1, Qwen3, Mistral), self-reported confidence inversel…