Extract the claims

An answer can feel convincing because it is well written. First strip away the style. Highlight concrete claims: names, dates, numbers, quotations, product availability and causal statements. Put them in a short table. A sentence like “This model is available to all users and saves 40 percent of costs” contains at least two separate claims. Ask which of them would change a real decision if false. Verify those first.

Second, inspect the source behind each high-impact claim. Prefer a current primary document for product details, a published paper for research methods and an official record for a policy. An AI-generated citation may be fabricated or may point to a page that does not support the sentence. Open the material yourself. Check the date, the actual wording and whether the source is describing a test, an estimate or a promise.

Challenge the interpretation

Third, compare the strength of the answer with the strength of the evidence. A company benchmark is not independent proof of superiority. A preprint is not a clinical treatment. A pilot program is not general availability. Fourth, seek a relevant counterexample or a second perspective when stakes are high. This can reveal a missing limitation without requiring you to treat every claim as equally controversial.

Fifth, rewrite the answer in calibrated language. Replace “proves” with “reported” when appropriate, specify who measured a result and state what remains unknown. If you cannot verify a claim, remove it or label it clearly as unconfirmed. Keep an evidence note for future updates.

Match effort to risk

You need not audit every recipe idea like a medical decision. Spend more effort where a mistake could cost money, harm someone or damage trust. For health, legal or financial choices, consult qualified sources and people rather than relying on a chatbot summary. The most valuable habit is simple: treat AI as a fast draft and a starting map, never as its own independent witness.