Why AI makes things up

Hallucinations are confident errors: names, citations, dates, or steps that fit the shape of a true answer and are not true. They happen because the model is optimizing for plausible continuation, not for a lookup against the world.

Humans treat smooth sentences as a proxy for competence. Models are extremely good at smooth sentences. That mismatch is why a wrong biography can feel more trustworthy than a hedged one. The fix is procedural: ask for sources, then open the sources; ask for uncertainty, then see whether the model actually marks it.

Ask for three citations on a niche local-history question you already know. Check whether the titles exist. Then ask the model to label each claim as verified, inferred, or unknown. A literate user rewards the unknown label instead of punishing it.

A wrong joke is cheap. A wrong medical, legal, or 'do my homework' answer is not. WonderMind refuses to be a homework machine in part because an unchecked fluent answer is how students learn the wrong habit.