How Do You Know? The One Question to Ask Every AI Answer

A confident answer feels like a checked one, but confidence tells you nothing.

Did the AI look the answer up, or did it piece the answer together from scraps? AI sounds the same either way.

Say you asked about a tax deadline, or how long a medicine takes to leave your body. Back came one calm sentence with a date in it, and no hedging or "I think", just "The deadline is the 31st."

That could be right, but it could also be last year's date, or a date pieced together from two half-remembered rules. You can't hear which.

Three words let you hear the difference, and they work in any tool, on any topic.

If it isn't a fact, what is it?

Every answer is one of three things, and none of them comes with a label.

The first is a search, where the AI found a web page and told you what the page says. That's good, if the page is right and up to date.

The second is an estimate, which the AI built from things it does know. That could be how common your surname is, or what the rule was the last time it looked. An estimate is often useful, but it's never a real count.

The third is a guess, where the AI had nothing solid, so it filled the gap. Filling gaps smoothly is what AI was built to do.

That last kind is the one that bites, because on the screen a guess looks exactly like a search.

What do I type?

The three words are "How do you know?"

Type them under any answer that matters, or for a more thorough version, use this.

How do you know that? For each part of your answer, tell me where it came from. Was it a source you can link to, an estimate you built from other things, or your best guess? Show me the links where you have them, and say plainly where you don't.

The AI won't sulk, and most of the time it answers honestly.

Your question gives it permission to say "I'm not sure", which AI rarely says unless you ask. Expect something like "The deadline is from the HMRC page here. I'm estimating the penalty amount from last year's figures. The bit about extensions is my understanding rather than something I checked."

One answer has become three, each part with its own level of confidence, and now you know where to look.

What do I do with what it says?

Learn the last two rows well, because when you challenge AI, it will sometimes dig in and sometimes flip.

Neither move is the truth, and both mean the AI was guessing. So watch for a change, or a refusal to change, with no source attached.

Do I have to ask every time?

No. On a recipe, asking just wastes a Tuesday.

Ask when being wrong would cost you something, such as money, a deadline, or a decision about your health or someone else's. It could also be anything you're about to tell another person as true.

I use one test, which is whether I'd be embarrassed to say "the AI told me" if it turned out wrong. If I would, I ask.

Ask hardest when the AI sounds most certain. Think of an answer that came back instantly, in one clean sentence, with an exact number, because that has the exact shape of a confident guess.

With AI, hesitation is a good sign, and a smooth, easy answer is a warning.

Practise somewhere harmless first. Ask AI how rare your name is, then ask how it knows. It takes five minutes, and after that you'll read AI answers differently.

Can I trust what it tells me about its sources?

Not always, because now and then the AI says "from this page", and the page says something else, or the page doesn't exist. That's why the link itself is the check, and the AI saying there's a link is not.

Pages go out of date, too. An answer can be wrong with a real source behind it, or it can be right by accident.

Three words get you to the evidence, but judging that evidence is your job.

No setting makes the AI ask itself. You type the three words every time it matters, until they become a habit.

This week, pick one thing AI told you that you've since repeated to someone as true. Then ask the AI how it knows.