Trust is a calibrated forecast, not a feeling

An AI gives you a confident, well-written answer and you feel like trusting it. Stop and notice what just happened: the feeling came from the prose, not from the facts. Fluency is a writing style. A model can be smooth and wrong in the same sentence, and the smoothness tells you nothing about whether the claim holds.

Treat trust as a forecast instead. Pick one checkable claim in the output — a number, a name, a date, a step that either works or doesn't. Probe just that one. If it holds, nudge your estimate up a notch; if it cracks, drop it hard. The goal isn't certainty, it's a probability you keep updating as evidence comes in. The trick is keeping each check cheap enough that you actually keep doing it, instead of verifying once and then coasting.

This pairs with two habits. It's how Delegate the task, keep the judgment in practice — you weigh the output rather than defer to it. And the claims easiest to probe are the ones that commit to something, because A good explanation is hard to vary and so it's easy to test where it would fail.

See trust protocols.

Next: on your next AI answer, pick one claim and actually check it.