How to check an AI answer: five checks before you use it

Checking an AI answer is the work the tool does not do for you. A language model produces plausible text; it has no way of knowing that it is wrong, and no reason to stop and warn you. You are the one who triggers the check, and the only moment when it is still useful is before the answer goes out.

This guide describes five checks, in the order in which they cost the least. None of them requires technical skill. Together, they are enough to intercept the vast majority of exploitable errors.

Why is a confident answer not a correct answer?

The tone of an answer says nothing about its truth. A model writes what it “knows” and what it reconstructs in exactly the same way. That is precisely what makes the error dangerous: it does not look like an error.

The best-documented case is a legal one. In June 2023, a federal court in New York sanctioned lawyers who had filed a brief citing non-existent court decisions, produced by a generative AI tool. The text was perfectly written, the references perfectly formatted, and none of it existed.

The corollary is counter-intuitive: the more precise an answer is, the more it deserves a check. An article number, a date, a percentage, a named court ruling: these are the easiest elements to fabricate and the most costly to let through.

Check 1: open every source cited

A reference is verified by opening it, not by reading it. This is the fastest and most discriminating check: an invented source never survives a click.

Three cases arise:

• The link does not exist, or leads somewhere else: remove the reference outright.
• The link exists but does not say what it is made to say: this is the most frequent case, and the most discreet.
• The link exists and confirms the point: keep it in your document, with the date you consulted it.

If the answer cites no verifiable source, treat it as a draft, not as a result.

Check 2: isolate what is precise

Reread the answer and highlight everything that can be verified: figures, dates, proper names, titles of legal texts, thresholds, obligations. The rest (structure, tone, wording) does not need to be verified, only judged.

This separation changes the nature of the work. You are no longer rereading a whole page with uniform vigilance: you are checking fifteen identified elements. It is faster, and above all it is repeatable.

For a figure that moves (a regulatory threshold, a price, a timetable), always ask for the date and the source. A value without a date is not information.

Check 3: ask the question differently

Open a fresh conversation, without the previous history, and ask the same question worded differently. Then compare.

• What comes back identical in both answers is generally solid.
• What appears only once should be checked first.
• What changes completely signals that the model is reconstructing rather than retrieving.

The fresh conversation is essential: within the same thread, the model sees its previous answer and tends to confirm it. You would not test a hypothesis by repeating it to the same witness in front of their written statement.

Check 4: ask what weakens the answer

Models go along with you. Ask “confirm that this clause is correct” and you will get a confirmation. The counter-measure takes two moves.

1. Ask the question in the negative. “What is wrong with this text?”, “which three risks have I missed?”, “where would an auditor pull me up?”.
2. State the opposite of what you think, then see whether the opinion holds. If the model changes position without a new argument, its first answer was worth nothing.

General rule: never ask a tool to validate a decision that has already been made. Ask it to attack the decision.

A doubt expressed is often enough to trigger a correction: state it explicitly rather than starting over from scratch.

Check 5: have it read by someone who knows

The last check is not technical. A credible but false answer can only be spotted by someone who knows the subject: and that is where the work shifts.

Output volume goes up; proofreading becomes the bottleneck and the scarce skill. The most experienced people gain in value, because only they can see what is wrong with a flawless text. Organise yourself accordingly: decide who reads what before the answer leaves the team.

Key points

• The tool does not check itself: you are the one who triggers the check.
• Open every source cited. An invented reference does not survive a click.
• Check what is precise: figures, dates, references, thresholds.
• Ask the question again in a fresh conversation, then compare.
• Ask what weakens the answer, never for a confirmation.
• Have it read by someone who genuinely knows the subject.

When is checking mandatory?

Three situations allow no exception: anything that leaves the organisation, anything that commits you legally or financially, and anything that touches the safety of people. A letter to a customer, a quote, a contractual clause, a safety instruction: human review is the minimum.

The European framework points the same way. Regulation (EU) 2024/1689 on artificial intelligence, published in the Official Journal of the European Union on 12 July 2024, imposes transparency and human-oversight obligations that grow stronger as the use becomes more sensitive. On the security side, ANSSI, the French national cybersecurity agency, published its recommendations for a generative AI system in April 2024. Read them as a reminder: generative AI is a component of the information system, not an oracle sitting beside it.

Where should you start in your job?

Start with the two notions that explain most errors, described in our guide AI vocabulary in four words. Continue with the question of data, covered in the guide what data can you share with an AI.

To put these reflexes into practice on cases from your own job, the course Generative AI at work and the artificial intelligence topic turn them into graded exercises.