Confidence calibration — the gap between how certain you feel and how accurate you actually are — is one of the most studied problems in judgment research. Philip Tetlock's decades of forecasting work found that the most dangerous predictors weren't the ones who said 'I don't know' too often; they were the ones whose certainty was immune to incoming evidence. Now consider what AI interfaces are quietly doing to this calibration. When a language model delivers an answer in the same confident, even prose whether it's reporting a well-documented historical fact or confabulating a plausible-sounding fiction, it trains your confidence-detection system on a broken signal. The Victorian logician Augustus De Morgan wrote about what he called the 'paradox of the unknown quantity' — that quantities we cannot measure tend to get treated as zero. The AI version of this is subtler: certainty you cannot see gets treated as present. The practical discipline here is not skepticism about AI output in general, but something more targeted — rebuilding the habit of asking 'what would a wrong version of this answer look like?' before you accept the right one. Because that question is exactly what fluent, unhesitating prose is optimized to make you skip.

In the last week, which AI-generated claim did you accept most quickly — and what would it have cost you to test it first?

Drawing from Philosophy of probability and judgment / Tetlock's forecasting research — Philip Tetlock (Superforecasting: The Art and Science of Prediction, 2015) with Augustus De Morgan (Formal Logic, 1847)