Fact-Checking What an AI Told You

Checking everything is not a strategy anybody sustains. Checking the three classes of claim that are wrong most often takes about five minutes and catches the errors that cost you something.

The five-minute routine

1. Underline the load-bearing claims. Not every sentence — the ones you would repeat to somebody else, or act on. Usually three to six in a long answer.
2. Check every citation exists before checking whether it says what was claimed. Search the exact title. Thirty seconds each, and this is the highest-yield check there is.
3. Check every proper noun paired with a specific fact. A person’s job title, a company’s founding year, which organisation published what. Names attract confident errors.
4. Recompute one number by hand. If there is arithmetic, do one step of it yourself. Errors in this class are rarely isolated.
5. Ask what would have to be true. For the central claim, name the evidence that would support it and check whether that evidence is what you were shown, or a plausible substitute.

That is the whole routine. What follows is why each step is there and how to do it quickly.

Which claims are actually worth checking

Error rates are not uniform across kinds of claim, and knowing the shape of the distribution is what makes checking affordable. Roughly, from most to least dangerous:

The mechanism behind the pattern is worth understanding once, because it predicts new cases: a model produces the most likely continuation, and there is no step anywhere in it that checks whether something is true. Fluency and accuracy are produced by the same process, so one is not evidence of the other.

Citations that do not exist

A fabricated reference has the right author names for the field, a title that sounds like a real paper in that literature, a plausible journal and a plausible year. It is undetectable by reading. It is trivially detectable by searching.

1. Search the exact title in quotation marks. If nothing comes back, it does not exist. Do not accept a similar-looking paper as confirmation — that is the trap, because a real adjacent paper usually does exist.
2. If it exists, check that it says what was claimed. A real source attached to a claim it does not support is the harder version of this error and it is common.
3. For legal, medical or regulatory citations, find the primary text. Section and article numbers are exactly the kind of predictable format that gets invented confidently.

Asking the model for its sources afterwards does not help. It will produce sources with the same machinery it produced the claim with. The only thing that works is looking.

Numbers, names and dates

A specific number attached to a general claim is the highest-risk sentence pattern in ordinary use: “studies show around 40% of organisations report…”. Where the figure is real, there is a named study behind it. Ask which one, then check that it exists and measured what the sentence says.

For arithmetic, recompute rather than re-ask. Ask the same model to check its own sum and you often get agreement, because the second pass is drawn from the same distribution as the first. A calculator, or a spreadsheet formula, is a genuinely independent check. That principle — the check must not come from the thing being checked — is the one to remember when the specifics here go out of date.

Dates are quietly unreliable in a specific way: a model has no dependable sense of what today is unless told, and no precise knowledge of where its own information stops. “As of 2026” in an answer is a sentence pattern, not a claim about currency.

Checking a claim you cannot look up

Some claims have no source to check: an estimate, a mechanism, an explanation of why something happens. You are not helpless, but the checks are different in kind — they test coherence rather than truth, so they can only ever raise suspicion.

• Check the order of magnitude yourself. Most wrong quantitative claims are wrong by a factor rather than a fraction. Build the number from two things you already know and see whether the answer is in the same neighbourhood. This is a two-minute exercise and it catches a surprising amount.
• Ask who would know. If a claim were true, some organisation would have measured it: a regulator, a statistical agency, a trade body, a professional register. If you cannot name one who would, the claim may be unmeasurable, and a specific figure for an unmeasurable thing is a warning rather than a finding.
• Test it against a case you know. Apply the general claim to a specific instance from your own experience. A rule that fails on the one case you can check is not a rule you should repeat.
• Ask for the mechanism, then check the mechanism. “Why is that true?” often produces a chain that contains one checkable link. Verify that link rather than the conclusion.
• Ask what would make it false. A claim whose stated falsifying condition is vague or absent is usually not a claim about the world.

Where none of these resolves it, the honest position is that you do not know, and the correct action is either to find one real source or to stop repeating the claim. That is not a counsel of perfection: passing on an unverifiable figure is how most of the bad numbers in circulation got there.

The checks that do not work

• “Are you sure?” This reliably produces either an apology and a changed answer, or reassurance, and neither correlates with the truth. Asking it after a correct answer will frequently talk the model out of a right answer.
• Asking for a confidence score. The number is generated the same way as the rest of the text. It is not a measurement of anything.
• Asking the same model in a new chat. Better than nothing, since it removes the influence of the conversation, but two samples from one distribution agreeing is weak evidence. A differently-worded question can change the answer without either version being more reliable.
• Trusting a linked source without opening it. A link can be real, live, and about something else entirely.

Prompting for checkability up front

Cheaper than checking is asking in a form that makes checking fast.

Answer the question. Then, before anything else, add a block
headed CHECK THIS:

- every claim in your answer that you are not confident about
- every number, and where it would come from
- every name or citation, marked EXISTS if you are confident the
  source is real and UNVERIFIED if you are not
- anything that depends on information after your training data

If you do not know something, write "I do not know" in the answer
itself rather than in the check block.

This is not a reliability measurement and should not be read as one. What it does is concentrate the risky material in one place so your five minutes go to the right lines. Instructions that make abstention acceptable do measurably change behaviour — see prompting a model to say it does not know.

When checking is not enough

Some answers should not be used even after they check out, because the cost of the error you did not catch is not symmetric with the time you saved. Medication doses, legal deadlines, tax positions, structural calculations, anything where being wrong is irreversible. The list and the reasoning are in the tasks where AI is the wrong choice.

There is also a limit to what checking can do for an answer that is wrong by omission. Verification tests the claims that are present. It cannot tell you about the consideration that never came up, and for anything consequential that is the more common failure. Reading one real source on the subject remains the only fix.