AI Is Getting Better at Knowing What It Doesn’t Know

One of the most persistent criticisms of large language models has been that they are confidently wrong — that they assert false information with the same linguistic confidence as true information, making it difficult to calibrate how much to trust any given output.

This is becoming less true. The calibration of frontier models — the alignment between their expressed confidence and their actual accuracy — has improved substantially in the past two years, driven by training techniques specifically designed to improve epistemic accuracy.

The improvement is measurable. On factual question-answering benchmarks that include a “I don’t know” option, frontier models in 2026 use that option more appropriately than their 2023 predecessors, and are more likely to express uncertainty on questions where their accuracy is genuinely lower.

The improvement is also incomplete. Models remain overconfident on questions at the edge of their training data, on recently updated information, and on tasks that require integrating multiple uncertain estimates. The hallucination problem has not been solved. It has been reduced, and the reduction is meaningful, and the work continues.