A good AI answer is clear, complete and confident. That is exactly why it is hard to check. A wrong date, a misquoted rule or an invented figure can sit inside a paragraph that reads perfectly well.
Checking an answer against an official source is a decision method, not a guarantee. It tells you which parts you can rely on, which parts are wrong, and which parts you cannot settle yet.
Step 1: Isolate the claims
Copy the answer and split it into single statements. One sentence can hold two claims: "The right of withdrawal is 14 days, and it applies to all online purchases" is two claims, and only one of them is probably true.
Skip the wording and keep the assertions: a number, a date, a rule, a name, a cause, a recommendation.
Step 2: Keep only the verifiable claims
Some statements cannot be checked against a source: opinions, advice on how to weigh options, general framing. Mark them as opinions and move on.
What is left is the list to check. For most answers it is short: three to ten claims.
Step 3: Find the primary source
A primary source is the document that creates the fact: the law text, the regulator's publication, the official statistics, the manufacturer's documentation, the scientific paper itself. A news article or an AI answer that quotes that document is not a primary source.
Look for the source that the claim would have to match. For a rule, that is the text of the rule. For a figure, that is the agency that publishes it.
Step 4: Compare, claim by claim
Read the source and compare it with the claim. Check four things:
- the number, date or name itself;
- the scope: who it applies to, where, and since when;
- the exceptions and conditions;
- whether the source is current.
Step 5: Conclude with one of three verdicts
Give each claim one status:
- Confirmed: the primary source says the same thing, with the same scope.
- Contradicted: the primary source says something different, or the claim is outside its scope.
- Not verifiable with a sufficient source: you cannot find a primary source that settles it, or the source is ambiguous.
Keep the third status. An absence of source is not proof that a claim is false. It means you do not yet know, and the answer should not be used for that point without more checking.
A worked example
An AI answer says: "In France, you can cancel an online purchase within 14 days without giving a reason."
- Claims: a period of 14 days; the right applies to online purchases; no reason is needed.
- Verifiable: all three are rules, so all three can be checked.
- Primary source: the consumer code and the EU directive on consumer rights, both public, plus the official consumer information published by the French administration.
- Comparison: the 14-day period matches. The right applies to distance sales, which includes most online purchases. The text also lists exceptions, for example some custom-made goods and some perishable items, and it sets conditions on how the request is made.
- Verdict: the period is confirmed. The "all online purchases" part is partly contradicted by the exceptions. The answer is right on the rule and too broad on its scope.
The lesson is not that the AI was wrong. It is that a correct rule can still be stated too broadly, and that only the source shows where the limits sit.
The same routine applies to medical, legal and financial claims. The stakes are higher, so the sources must be more recent and more specific, and "not verifiable" should lead you to an expert rather than to a guess.
Limits of this method
- A primary source can be out of date. A rule may have changed after the source was written, so check the date of the text.
- A source can be ambiguous, or you can misread it. If two careful readers would interpret it differently, mark the claim as not verifiable.
- A source may not cover your case. A national rule does not always apply to a cross-border purchase or a specific contract.
- Checking takes time. For a casual question, one check on the claim that matters most is enough.
When the models also disagree
Asking the same question to several AI models can show you which claims are unstable. A claim that several models give in different forms is a good candidate for checking. A claim on which they agree is not thereby true: they can share the same error. The method above still decides.
If you want to see where independent answers diverge before you check them, Satcove asks several models the same question and shows the points of agreement and disagreement, so you know which claims to verify first. It does not replace the primary source, and it does not decide what is true for you.
For the method behind consensus and its limits, read the guide to verifying AI answers and the definition of AI consensus.