A good explanation is hard to vary

The physicist David Deutsch put it sharply: a good explanation is hard to vary. Every part does load-bearing work, so you can't change a detail without wrecking the account. A bad explanation is the opposite — it's plausible, fluent, and would fit any outcome just as comfortably. "Sales rose because the market shifted" explains a rise, and you could tell the identical story about a fall. It commits to nothing, so it predicts nothing.

This is exactly the kind of text an AI is good at producing: confident, post-hoc prose that drapes over whatever happened. The defence is to demand the parts that could be wrong — the specific numbers, mechanisms, and dependencies — and then check them. That's where Show the data; don't just assert it, because the evidence is what an easy-to-vary story quietly skips. And it's why Structure beats freeform text for working with AI: a schema forces the commitments into the open instead of letting them hide in flow.

A claim that can't be wrong can't be trusted either — so it's worth asking what the trust is resting on. One answer is to treat Trust is a calibrated forecast, not a feeling as a forecast rather than a feeling. See whether it holds up against the frameworks for explanation.

Next, then: the evidence is what an easy-to-vary story skips — so Show the data; don't just assert it.