Can models express uncertainty instead of just answering?

Even on the simplest setting — factoid QA with clear ground truth and no external tools — frontier models still hallucinate. The paper's diagnosis is that most factuality gains have come from expanding the model's knowledge boundary (encoding more facts) rather than improving awareness of that boundary (distinguishing known from unknown). It conjectures the latter is inherently hard: models may lack the discriminative power to perfectly separate truths from errors, creating an unavoidable tradeoff between eliminating hallucination and preserving utility.

That tradeoff dissolves under a reframing. If hallucination is understood as confident error — incorrect information delivered without appropriate qualification — then a third path opens beyond the answer-or-abstain dichotomy: expressing uncertainty. The proposal is faithful uncertainty: aligning the model's linguistic uncertainty with its intrinsic uncertainty. This is one facet of metacognition — being aware of one's own uncertainty and acting on it.

The framing's reach is what makes it post-worthy. Faithful uncertainty becomes the control layer for robust agentic tool use, and it is fundamentally a form of honesty — accurately representing one's epistemic state rather than projecting false confidence — which connects it to AI safety. It also enables appropriate human oversight: a model that expresses calibrated doubt invites users to verify and exercise judgment. Realizing it requires shifts on both sides — benchmarks that reward calibrated uncertainty rather than only accuracy, and users who expect and can interpret it. This complicates Does reasoning fine-tuning make models worse at declining to answer?: faithful uncertainty is the richer target that pure abstention only crudely approximates.