An AI answer can be fluent, specific, and wrong at the same time. The practical response is not to distrust every sentence equally. It is to verify in proportion to the harm a mistake could cause.

The National Institute of Standards and Technology calls confidently presented false content confabulation in its Generative AI Profile. NIST also warns that generated citations and reasoning can themselves be fabricated. For everyday use, that observation becomes a short habit: identify the decision, isolate the claims, and check the highest-risk ones against primary evidence.

This workflow is designed for ordinary research and workplace questions. It is not a substitute for a doctor, lawyer, accountant, security professional, or another qualified expert when the consequences are serious.

The five-minute version

If you are short on time, use this sequence:

  1. Write down what decision the answer will influence.
  2. Highlight names, dates, numbers, quotes, rules, and links.
  3. Open the best primary source for the most important claim.
  4. Check whether the source supports the exact wording, including conditions and dates.
  5. Record what remains uncertain before you act or share the answer.

Five minutes will not validate every complex answer. It will often reveal the most dangerous failure: a confident conclusion resting on a source that is missing, irrelevant, or out of date.

1. Name the decision

Before checking the answer, ask what you plan to do with it.

If the output is a brainstorming list, a minor error may cost a minute. If it influences security settings, a purchase, medication, taxes, legal rights, employment, or public claims, the verification bar should be much higher—and a qualified professional may be necessary.

A useful prompt to yourself is: What happens if this answer is wrong? The answer determines how much checking is reasonable.

Use three rough risk levels:

  • Low stakes: brainstorming, rephrasing, meal ideas, or a disposable draft. A quick reasonableness check may be enough.
  • Medium stakes: product comparisons, travel planning, software configuration, or material you will publish. Verify the facts that affect cost, compatibility, timing, or reputation.
  • High stakes: health, legal rights, taxes, finance, safety, cybersecurity, employment, or decisions about another person. Treat the AI output as a lead, not an authority, and involve an appropriate professional or official channel.

2. Turn prose into checkable claims

AI systems often mix facts, interpretation, and advice in one smooth paragraph. Split the answer into smaller statements:

  • factual claims that can be confirmed;
  • calculations that can be repeated;
  • recommendations that depend on priorities;
  • assumptions that may not match your situation;
  • predictions that cannot yet be verified.

This matters because “the paragraph sounds right” is not a test. A claim such as “this feature launched in March” has a source. “This is the best option” needs criteria.

Numbers deserve their own pass. Recalculate totals and percentages from the original figures. Check whether a rate is monthly or annual, whether a price includes tax, and whether a statistic describes the same population and time period as the conclusion. A correct-looking calculation can still be based on the wrong inputs.

3. Ask for sources—then open them

Requesting citations is useful only if you inspect them. A citation can be real but irrelevant, outdated, or weaker than the sentence it is supposed to support. In some cases, a model may produce a title or URL that does not exist.

Prefer sources closest to the fact:

  1. an official specification, government publication, court document, or vendor support page;
  2. original research or a dataset;
  3. reputable reporting that links to the underlying evidence;
  4. commentary and summaries for context, not as the sole proof of a consequential claim.

Check the publication or revision date. Software instructions and policies can change quickly.

Do not treat a list of links as evidence by itself. For each important link, ask:

  • Does the page exist and identify a responsible publisher?
  • Does it actually contain the claim?
  • Is the claim current for the relevant version, jurisdiction, or date?
  • Is the source reporting direct evidence, or merely repeating another summary?
  • Does the source describe a product’s existence, or does it also support claims about performance and reliability?

If an answer quotes a source, search the source for a distinctive phrase. A quote that cannot be found may be paraphrased badly, taken out of context, or invented.

4. Look for the missing condition

Many misleading answers are not completely false; they omit a condition. Watch for phrases such as “always,” “guaranteed,” “all devices,” or “works with.” Then ask what version, region, plan, device, account type, or date the claim assumes.

Also look for evidence boundaries. A vendor announcement can confirm that a feature exists. It does not independently prove reliability, performance, or how well it works for your use case.

For instructions, verify the starting state as well as the steps. A command written for Windows may not apply to macOS. Account menus can differ by plan, region, or administrator policy. A technically accurate procedure can still be wrong for the system in front of you.

5. Run a contradiction check

Before accepting the conclusion, try to disprove it. Search an authoritative source using a neutral query, not the AI’s exact wording. Compare at least two independent sources when a claim is consequential or disputed.

For a recommendation, ask the model for the strongest reason not to follow it, then verify that reason separately. This is not because the second answer is more trustworthy; it is a way to uncover assumptions that the first answer may have hidden.

If two reliable sources disagree, do not average them. Check whether they address different dates, definitions, populations, product versions, or jurisdictions. If the conflict remains unresolved, report the disagreement.

6. Make the result auditable

For anything you may revisit, save a short evidence note:

  • the question you were answering;
  • the conclusion you reached;
  • the primary links you opened;
  • the date you checked them;
  • any unresolved uncertainty.

This is faster than rebuilding the reasoning later, and it makes updates possible when the underlying facts change.

A compact evidence note might read: “Checked the vendor support page and the current specification on October 21, 2026. The feature exists on paid plans, but the documentation does not support the claimed performance improvement.” That separates what the evidence shows from what the AI inferred.

Common warning signs

Pause and verify more carefully when an answer includes:

  • an exact quote without a page or section reference;
  • a citation whose title does not match the linked page;
  • precise statistics with no dataset, date, or population;
  • claims about a law or policy without a jurisdiction and effective date;
  • instructions that require disabling a security control;
  • absolute language such as “always,” “never,” or “guaranteed”;
  • a recommendation that does not state the criteria or alternatives;
  • claims of firsthand testing that the system could not have performed.

These signals do not prove the answer is wrong. They identify where verification has the highest payoff.

A practical stop rule

Verification can expand forever. Stop when the strength of the evidence matches the consequence of the decision.

For low-stakes work, one authoritative source and a quick plausibility check may be enough. For medium-stakes work, check the decision-changing claims and record the sources. For high-stakes work, use multiple independent sources and the appropriate expert. If the evidence is missing, say unknown rather than turning the model’s confidence into your own.

The goal is not to prove that an AI tool is trustworthy in general. It is to decide whether this particular answer is supported well enough for this particular use.