A Fact-Checking Checklist for Claude's Answers

Use this checklist before you repeat, publish, or act on a specific claim, number, or citation that Claude gave you.

It turns the Discernment and Diligence skills from the AI Fluency Framework into a concrete, repeatable routine.

How to Use This Checklist

• Not every item applies to every answer; scale your effort to the stakes of what you're doing.
• Work through the tiers in order: Tier 1 first, since it screens out the biggest risks fastest.
• For a low-stakes, casual question, Tier 1 alone is often enough.
• For anything you'll publish, ship, or hand to someone else, work through all three tiers.
• Keep a running note of what you verified and how, especially for longer documents with many claims.

Tier 1: Triage the Answer (do this first, always)

1. Identify the load-bearing claims: find the one to three specific facts, numbers, or claims your decision most depends on.
   • These are the ones worth the most scrutiny; you don't need to verify every sentence.
2. Separate fact from framing: mark which sentences are verifiable facts and which are Claude's interpretation, summary, or opinion.
   • Interpretation isn't wrong by default, but it shouldn't be treated as a verified fact.
3. Check for hedging language: note any place Claude used words like "likely," "I believe," or "I'm not fully certain."
   • Treat a hedge as a direct signal to verify that specific point before relying on it.
4. Scan for suspiciously precise numbers: flag any statistic given to an oddly exact decimal or figure with no stated source.
   • Precise-sounding numbers can be generated the same way as any other text and are not automatically accurate.
5. Note any citation or named source: list every study, article, author, or publication Claude referenced.
   • Each one goes into Tier 2 for independent confirmation.

Tier 2: Verify Specific Claims

6. Search for the claim independently: look up the fact or number using a search engine, a primary source, or a domain-specific reference.
   • Don't stop at the first result that happens to agree; check that it's actually a credible, relevant source.
7. Confirm the citation actually exists: for any named study, article, or publication, verify it is real before trusting the claim attributed to it.
   • A fabricated citation is one of the more common hallucination patterns, so this step is not optional for anything you'll cite yourself.
8. Confirm the source says what was claimed: open the source and check that the specific claim actually appears in it, not just that the source exists.
   • Sources are sometimes real but the claim attributed to them is not accurately represented.
9. Check the date and recency of the claim: confirm whether the fact is still current, especially for fast-moving topics.
   • An accurate fact from an outdated source can still mislead if presented as current.
10. Verify numbers against their original unit and context: confirm a statistic isn't being applied to the wrong population, time period, or scope.
    • A correct number used in the wrong context produces a misleading claim even though the number itself is accurate.
11. Cross-check with a second independent source when the claim is significant: don't rely on a single confirming source for a high-stakes claim.
    • Two independent sources agreeing is stronger evidence than one, especially for contested or unusual claims.

Tier 3: Review the Whole Answer

12. Re-read for one-sided framing: check whether the answer favors one interpretation without acknowledging reasonable alternatives.
    • This overlaps with watching for bias, not just factual accuracy.
13. Check internal consistency: confirm the answer doesn't contradict itself between its introduction, body, and conclusion.
    • Long answers can drift and produce quietly conflicting claims.
14. Re-verify assumptions carried over from earlier in the conversation: for long threads, confirm an early detail Claude is still relying on hasn't been misremembered.
    • Verification needs to be renewed across a conversation, not performed once at the start.
15. Decide what still needs a disclaimer: for anything you didn't fully verify, note that explicitly rather than presenting it with full confidence.
    • Passing along unverified claims as settled fact shifts risk onto whoever reads your work next.

Applying the Checklist in Order

• Tier 1 (fast, always): takes a minute or two and screens the answer for where your attention is actually needed.
• Tier 2 (targeted): apply only to the load-bearing claims identified in Tier 1, not the whole answer.
• Tier 3 (holistic): reserved for anything going to publication, production, or another person, since it checks the answer as a whole rather than claim by claim.

FAQs

Related

• Why Claude Sometimes Hallucinates - the mechanism behind the confident-but-wrong claims this checklist catches.
• Bias in AI Outputs: What to Watch For - a related check for framing, not just factual accuracy.
• Signs an AI Output Needs a Second Look - red flags that often trigger the need for this checklist.
• Discernment & Diligence Basics - starter habits this checklist builds on.

Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.