How to tell if an AI answer is true in 30 seconds

This is the one I never skip. Take something ordinary. Say you’ve pasted in your phone contract and asked for a plain summary. Before you read a word of that summary, make the AI tell you what it’s looking at: which provider, which plan, the monthly price printed at the top. “Which provider is this, and what am I paying a month?”

This sounds too obvious to bother with. It isn’t. An AI that’s muddled about which thing you mean will cheerfully answer about a different one and never flag the swap, like a waiter who brings the wrong table’s order and reads it out to you with total confidence. If it says £29 when the page in front of you says £39, stop there. It’s describing someone else’s contract in a very reassuring voice.

I added this half later, once I’d noticed the first question alone wasn’t enough. Now make it hand you something you can verify yourself, fast: one date, one number, one name you can hold against the document. A made-up number gives itself away the second you check it. A right answer about the wrong thing never does: every fact is correct, neatly laid out, and about a contract you’ve never seen in your life.

Here’s the rule I’m strictest about with myself. Not the bad line. The lot. If it couldn’t tell you what it was looking at, nothing it said next was about your problem; it was a tidy, confident essay answering a question you never asked. The good-looking paragraphs are the part that nearly fooled you.

Two questions won’t catch all nine ways AI gets things wrong. They catch the expensive ones fastest, and the rest of the checks are on the Guardrails page.

It works the same on anything that matters: a contract summary, a letter from your doctor, a holiday booking, an email from your child’s school. Name the thing. Check one fact. Then decide whether to believe a word of it.

Next: the whole method, in four questions → The whole method, in four questions