AI output verification checklist

Work through it in six passes: identify the claims that need checking, trace each one to where the information actually came from, verify the important ones against credible sources, challenge the reasoning connecting them, correct or qualify what does not hold, then review the revised output again before you use it.

You cannot tell from how the writing sounds — fabricated content reads exactly like accurate content. What you can do is check the claim. Unusually specific statistics with no source, citations you cannot locate, quotes with no traceable original and confident detail about obscure topics are all worth investigating. They are warning signs, not proof.

Open them. Confirm the source exists, that the title, author and publication details are right, and that the cited page actually contains the information attributed to it. Then check it supports the strength of the claim — a citation can be entirely real and still not say what the output claims it says.

It can help. AI is useful for identifying which statements are checkable claims, spotting internal contradictions, and drafting the verification work. But asking the same model to confirm its own answer is accurate is not independent verification — the check inherits whatever produced the error. Use it to organise the work, not to sign it off.

Anything factual that a decision or a reader could rely on: statistics, dates, names, quotes, citations, product capabilities, causal claims, legal or regulatory statements, and comparisons. Opinions, suggestions, structure and phrasing do not need verifying.

No. Match verification to risk. A brainstorm needs a light pass; a client deliverable or a business decision deserves a deep one. The higher the consequence of an error, the stronger the evidence and review should be.

Yes. Nothing here is model-specific. The checklist covers failure modes common to current AI systems, so it applies to any of them — and to the next ones.