Checklist: verify an AI answer before trusting it
An AI can be confidently wrong. Here is the 6-point checklist to apply before forwarding, publishing or acting on an AI-generated answer.
A wrong AI answer has exactly the same tone, the same fluency and the same confidence as a correct one. That is the whole problem with hallucinations: nothing in the form betrays the substance. The only reliable protection is a checklist applied before forwarding, publishing or acting. Here it is, in six points.
1. Identify the risk level of the answer
Not every answer deserves the same level of vigilance: the first step is to assess what is at stake before deciding how much verification is needed. Brainstorming ideas or rephrasing a tone has almost no consequence if the AI gets it wrong. A figure sent to a client, a date cited in a report, or a claim published online has a real one.
Three signals should trigger systematic verification:
- the answer contains a figure, a date, a proper name or a precise reference;
- the topic is recent, niche or poorly documented;
- the answer will serve as the basis for a decision, a send or a publication.
As soon as a single one of these signals is present, move to the following steps. Otherwise, a light check (step 2) is usually enough.
2. Reread with a critical, not a complacent, eye
The most useful reread is the one that actively looks for the error, not the one that confirms a general positive impression. Ask yourself: would this sentence be true if I checked it right now? A well-written answer can hide a shaky claim in the middle of an otherwise correct paragraph — it is even one of the most frequent traps, because the quality of the form wrongly reassures about the substance.
3. Demand and open the sources
A source cited by an AI is only proof if you open it. Systematically ask “where does this information come from” or “cite your sources” for any precise fact, then click the provided link. An invented source does not survive this simple test: the link is dead, the article does not say what is reported, or the document simply does not exist.
This reflex is especially important if you used an assistant to synthesise a document you have not read in full yourself: see our guide to summarising a long document without betraying it.
4. Cross-check with a second source, human if possible
For anything that truly matters — a financial figure, a legal reference, a medical fact, an exact quotation — a second, independent source remains indispensable. A quick search on a classic engine, an official document, or the opinion of a knowledgeable colleague on the subject is often enough to confirm or refute in a few minutes what the AI claims.
Do not ask the same AI for confirmation in the same conversation: it tends to stay consistent with its first answer rather than genuinely reconsider it.
5. Spot the topics with structural risk
Some types of requests produce more errors than others, regardless of the tool used. Caution must be reinforced on:
- calculations and statistics, where a reasoning error easily goes unnoticed in an otherwise well-written answer;
- legal or regulatory references, which change and vary by jurisdiction;
- recent events, on which the model may only have partial or dated information;
- exact quotations (author, work, date), one of the most frequent cases of plausible invention.
6. Document what is verified — and what is not
On content meant to be forwarded or published, keep a record of what you personally verified and what remains an unconfirmed AI claim. This distinction prevents you from later forwarding, without thinking twice, a piece of information you had in fact never validated yourself.
The decision table
| Type of answer | Verification level | Action |
|---|---|---|
| Rephrasing, brainstorm, draft | Light | A quick reread is enough |
| Summary of a document you own | Medium | Compare the key points to the original |
| Figure, date, proper name, quotation | High | Open the source, cross-check with a second |
| Legal, regulatory, medical reference | Maximum | A second human source is mandatory |
| Content meant for publication or external send | Maximum | All the steps above, before distribution |
Key takeaway
An AI that is wrong never signals it by tone. The only effective protection is verification proportionate to the risk: light for an internal draft, systematic as soon as a figure, a date or a precise reference is involved, and always backed by a second human source before any publication or decision. This reflex follows the same logic as the confidentiality precautions to adopt at work: treat AI as a powerful tool, not as an authority.
Sources
Frequently asked questions
Should you verify every AI answer, even the simplest ones?
No. The level of verification should be proportionate to the risk: a rephrasing or a brainstorm does not demand the same rigour as a figure, a date or a legal reference that shapes your decision.
How can you quickly spot a risky answer?
Three signals: it contains a precise figure, date or name; it concerns a recent or niche topic; it will serve as the basis for a decision, a send or a publication. As soon as one of these three is present, verify before acting.
Is an AI that cites its sources automatically reliable?
No. A cited source can be invented or distorted. The only reliable test is to open the link and check that the content really matches what the AI reported from it.
What if I cannot verify a piece of information another way?
Do not treat it as an established fact. Reframe it as a hypothesis to confirm, or explicitly ask the AI for its level of certainty and the limits of what it is claiming.
Does verification take as long as doing the work yourself?
No, in the vast majority of cases. Checking a figure or a source takes a few minutes; it is far faster than producing the information from scratch, and much cheaper than a published or forwarded error.