Definition
Plain language
Whether a system's willingness to act matches how much it actually knows.
As stated in the literature
Alignment between a model's decision to commit to an answer and the genuine resolvability of the question, as opposed to belief calibration, which only concerns stated probabilities.
Also called: belief calibration
Why it matters: Systems that state low confidence but still answer unanswerable questions push users into acting on noise, so matching willingness-to-answer with actual knowability is what keeps hedged guesses from being treated as findings.
For example, a well-action-calibrated assistant asked for tomorrow's exact lottery numbers would decline rather than confidently produce six digits, even if it hedges by saying it is only 20% sure.
Heard on the show
“Belief calibration and action calibration aren’t the same thing.”Episode 251 — When a Fake Dashboard Makes an AI Agent Just as Confident