Confidence is not the same as being right
One of the first things my AI coursework made me take seriously was calibration: whether a system's stated confidence matches how often it is actually correct. It sounds like a technical footnote. It turned out to be the most personal idea on the syllabus.
A confident answer and a correct answer look identical until you keep score.
Two different questions
Accuracy asks how often you are right. Calibration asks whether you know how often you are right. A classifier can score well on the first and badly on the second, predicting with 99% confidence on cases where it is really guessing. The number looks like knowledge. It is closer to tone of voice.
The fix is rarely more cleverness. It is usually more honest measurement: hold out data the model has never seen, bucket its predictions by confidence, and check whether the 70% bucket is really right seventy percent of the time. Reliability diagrams are humbling because they are so simple.
The version of this that lives in meetings
In three years of shipping web and mobile software, the most expensive mistakes I saw were rarely from not knowing something. They came from someone, sometimes me, saying "that will take two days" with the confidence of someone who had already done it, when they had only done something that looked similar.
Estimation is a calibration problem wearing a project-management costume. The useful habit is not to stop estimating but to keep score: write the guess down, compare it later, and notice which kinds of work you are consistently optimistic about.
Saying "I don't know" with a number attached
Large language models have made this question public. They produce fluent answers whether or not the underlying evidence is strong, and fluency is very persuasive. Building systems that can signal uncertainty, abstain, or show their sources is not a politeness feature; it is what makes them safe to rely on.
I am trying to hold myself to the same standard. Not less confident, just better calibrated: "I'm fairly sure", "I'd bet on it", and "I have no idea" should mean different things, and people should be able to learn which one I mean.