Detect hallucinations by turning an answer into individual claims and checking each against current, authoritative evidence. Mark claims as supported, contradicted, or not established. Correct the draft from the evidence, not from another unsupported model answer, and preserve uncertainty where the available source is incomplete.
Who this is for: Researchers, writers, analysts, and workplace reviewers who use AI drafts containing factual or source-dependent claims.
- Check atomic claims instead of asking whether an entire polished answer feels correct.
- Use source quality and direct support, not agreement among generated responses, as the standard.
- Remove unsupported specificity rather than replacing it with a plausible guess.
Recognize where hallucinations enter
Hallucination is a practical label for output presented as factual even though it is unsupported, false, or fabricated. Common forms include invented citations, merged people, wrong dates, nonexistent product features, and explanations that add causes absent from the evidence. The output may remain grammatical and internally consistent.
Risk rises when the prompt demands an answer despite missing information, asks about events outside available knowledge, or requests exact details from vague context. Summaries can also hallucinate by adding a conclusion that the source never makes. Plan verification based on consequence and claim type, not on how ordinary the prose appears.
Build a claim ledger
Split the response into the smallest claims that could be independently true or false. 'The policy started in May and reduced delays by 20 percent' contains at least two claims. Assign each a required source type and status: supported, contradicted, not established, or not checked. Quote the exact evidence when possible.
Prioritize names, numbers, dates, quotations, causal statements, legal or medical assertions, and descriptions of current software behavior. Low-impact connective language may need less effort. A ledger exposes mixed sentences where one supported fragment makes an unsupported fragment seem credible.
Choose authoritative evidence
Prefer primary documents, official records, standards, original research, and current product documentation appropriate to the claim. Search snippets and model-generated citations are leads, not evidence. Open the source, verify that it exists, confirm date and authorship, and read enough context to know whether it actually supports the wording.
For volatile topics, check the publication or update date and consult current official guidance. Two weak sites repeating the same unsourced statement do not create independent confirmation. When reliable sources disagree, describe the disagreement and its scope instead of forcing a single confident conclusion.
Correct and learn from the failure
Rewrite from the evidence ledger. Delete invented details, narrow overbroad claims, fix contradictions, and label remaining uncertainty. Do not preserve a fabricated citation merely because the underlying general point is true. The corrected draft should make its strongest claims no stronger than the sources allow.
Then diagnose the workflow. Provide better source material, permit abstention, require evidence locations, or separate research from drafting. Add the failed case to an evaluation set. Repeated hallucination is a system-design signal, not a proofreading inconvenience, especially when output reaches customers or decisions.
Audit an AI-written market brief
A brief claims a competitor launched a feature in March, gained 40 percent adoption, and discontinued an older plan.
- Split the sentence into launch, date, adoption, and discontinuation claims.
- Assign official release notes to the feature claims and a primary company report to the adoption figure.
- Open each cited page, verify identity and date, and quote the lines that directly support each claim.
- Mark the launch as supported, the month as contradicted, adoption as not established, and discontinuation as supported only in one region.
- Rewrite the brief with the corrected date, remove the percentage, narrow the plan statement, and record the case for future testing.
Claim verification ledger
Use one row per factual claim rather than one row per paragraph.
- Claim text and claim type: identity, number, date, quote, cause, capability, or interpretation.
- Required authority: official record, primary study, product documentation, or qualified expert review.
- Evidence location: source URL, date, section, page, and exact supporting passage.
- Status and correction: supported, contradicted, not established, or not checked, plus revised wording.
- Workflow lesson: missing input, forced answer, stale source, ambiguous task, or review failure.
Common mistakes
- Asking the same model whether its earlier answer was true and accepting its reassurance.
- Checking that a cited page exists without confirming that it supports the specific claim.
- Correcting one visible number while leaving unsupported causal language around it.
Try one
An AI answer cites a real NIST page but attributes a quotation that does not appear there. How should you classify and repair the claim?
Classify the quotation as unsupported or fabricated even though the page itself is legitimate. Check whether the source expresses a related idea in different words, then either paraphrase accurately with a precise reference or remove the claim. A complete answer records the citation failure and avoids inventing a replacement quotation. Evaluation should distinguish source existence from direct textual support.
Sources
- NIST AI Risk Management FrameworkNIST framework for governing, mapping, measuring, and managing AI risk.
- OpenAI safety best practicesOfficial recommendations for moderation, testing, and human oversight.