How to Check an AI Answer Before You Act on It

Direct answer

The short answer

This guide is for checking ordinary AI answers before you rely on them. It is not a substitute for a clinician, lawyer, accountant, engineer, or other qualified professional who can examine the evidence and accept responsibility for high-stakes advice.

Not every sentence deserves the same effort. A fictional name suggestion and a medication interaction do not carry the same consequence. Verification should be proportional to risk, recency, reversibility, and how much the claim affects your decision.

Triage

Verify by consequence, not by paragraph order

Highlight the smallest claim that would change what you do. “This contract can be cancelled within 14 days” matters more than the answer’s general explanation of consumer rights. “The recommended dose is safe with your medication” matters more than the surrounding description of side effects.

1. High consequenceHealth, law, money, safety, security, reputation, and irreversible actions require independent expert or authoritative confirmation.
2. Time sensitivePrices, policies, officeholders, software behavior, schedules, and regulations may have changed after the model’s knowledge or source date.
3. Precision dependentExact numbers, units, quotations, citations, and eligibility rules are easy to state plausibly and easy to get subtly wrong.
4. Low consequenceBrainstorming, wording options, and fictional ideation can be evaluated mainly by usefulness—unless they introduce factual claims.

Checklist

A seven-step AI answer verification checklist

1. Extract the claimsSeparate checkable facts from advice, opinion, assumptions, and generated examples.
2. Rank the riskIdentify which error would most change the decision or harm someone.
3. Choose the right authorityUse the law or regulator for rules, official product documentation for features, and original research for study findings.
4. Check scope and dateConfirm country, population, product version, plan, model, and publication or update date.
5. Trace the exact supportFind the sentence, table, dataset, or rule that supports the claim. Topic similarity is not evidence.
6. Reproduce the resultFor numbers, write down inputs, units, formula, assumptions, and rounding. Recalculate independently.
7. Record uncertaintyNote conflicting sources, missing evidence, and what additional fact would change the conclusion.

Source matching

What to check for different kinds of claims

• Medical: official public-health guidance, drug labeling, clinical guidelines, and a qualified clinician who knows the person’s history.
• Legal: current statutes, regulations, court or agency sources, correct jurisdiction, and a qualified lawyer for application to facts.
• Product: current official documentation, pricing, status pages, release notes, and the exact plan or version.
• Scientific: the original paper, methods, sample, outcome measured, limitations, retractions or corrections, and later evidence.
• Current event: event date rather than only publication date, direct statements or records, and multiple independent reports when facts remain disputed.
• Quotation: the original transcript or document, exact wording, speaker, date, and surrounding context.

An answer that links to a source is easier to audit, but a link alone does not establish support. Open it. A hallucinated citation may have an invented title, while a subtler failure uses a real source that never makes the claimed assertion.

Use the model carefully

Ask AI to expose uncertainty—not to certify itself

This prompt can reveal what needs checking, but it does not perform independent verification. Follow the links, inspect the originals, and reproduce important calculations outside the model. If the issue is high stakes, take both the question and the source material to a qualified person.

Good verification does not mean proving every sentence true. It means knowing which claims are supported, which remain uncertain, and whether the remaining uncertainty is acceptable for the decision you are about to make.