How to fact-check AI before you trust it

This is the most important habit on the whole site. AI can be confidently, smoothly, persuasively wrong, especially on specifics — and it sounds exactly as sure when it's inventing something as when it's right. A 30-second check is the entire difference between using AI as a superpower and getting publicly embarrassed by it.

Language models produce plausible, well-formed text, not verified facts. Most of the time plausible and true line up — which is why it's so useful. But on exact numbers, dates, names, quotes, citations, and legal or medical details, plausible and true can diverge, and the model has no internal alarm when they do. It will hand you a fake statistic with the same confidence as a real one. Knowing this isn't a reason to distrust AI; it's a reason to know exactly where to check.

Ask the model to split what it's confident about from what you should verify, and to flag any statistic, quote or source. Given explicit permission to admit doubt, it's surprisingly honest about which parts it's guessing on. That gives you a built-in risk map of its own answer, so you spend your verification effort only where it actually matters instead of treating every sentence as equally reliable.

For anything you'll publish or act on — work, money, health, a public post — open the actual source and confirm it really says what the AI claims, and cross-check important facts in a second tool or a primary source. The classic trap: it gives three real-looking links and one is a 404 it invented. If you'd pasted that 'source' into a report unchecked, the mistake would have been yours, not the AI's.

1. 1After an answer, ask it to separate confident claims from ones to verify.
2. 2For any stat, quote, or source, open it and confirm it actually says that.
3. 3Cross-check anything consequential in a second tool or primary source.
4. 4Do the check before you publish or act, not after.

Verification prompts