Verification workflow
How to fact-check ChatGPT: a claim-by-claim workflow
Do not ask only whether an answer feels right. Turn it into a list of propositions that evidence can support or reject.
Before you start: preserve the original answer
Copy the answer or share a stable transcript before prompting again. A revised response can silently change claims, making it hard to see what was corrected and what was merely rewritten.
Seven steps
- 01
1. Split the answer into atomic claims
One row per fact: who, what, when, where, number, requirement, cause, quotation, or recommendation premise. Separate compound sentences.
- 02
2. Rank the claims by consequence
Check first what affects money, health, safety, rights, deadlines, grades, production systems, or a public statement.
- 03
3. Run an adversarial second pass
Ask the model to name assumptions, ambiguities, stale facts, disputed points, and the three claims most likely to fail. This creates leads, not proof.
- 04
4. Find primary and authoritative sources
Use the actual law, standard, documentation, paper, dataset, filing, product page, schedule, or original recording where possible.
- 05
5. Inspect the source, not just the snippet
Confirm the exact sentence, context, date, version, jurisdiction, sample, and limitations. Search snippets and AI summaries can omit the part that changes the conclusion.
- 06
6. Recalculate and reproduce
Check arithmetic, units, filters, code, and transformations independently. For data claims, locate the numerator, denominator, and time window.
- 07
7. Publish a verification note
Mark each claim verified, contradicted, partly supported, or unresolved. Link the evidence and record the date you checked it.
A reusable verification table
| Claim | Why it matters | Evidence needed | Result |
|---|---|---|---|
| The policy changed on July 1 | Determines whether the planned action is allowed | Official policy and effective-date notice | Verified / contradicted / unresolved |
| The study found a 35% improvement | Supports the recommendation | Original paper, outcome definition, sample, and confidence interval | Verified / narrower than stated / unresolved |
| This API option exists in version 4 | Affects implementation | Versioned official documentation or source code | Verified / renamed / removed |
How to inspect citations
- Search the exact title, author, journal, case number, or URL. No result is a strong warning sign.
- Open the original and search for the quoted phrase. A near match is not necessarily the quotation claimed.
- Check whether the cited source is primary, a commentary, or another AI-generated summary.
- Read the conclusion and limitations, not only the abstract or headline.
- Confirm publication date, correction notices, retractions, archived versions, and whether newer guidance supersedes it.
Prompts that make fact-checking easier
- Convert your previous answer into atomic factual claims. Do not revise them yet.
- For each claim, state what primary source would be needed to verify it and what evidence would contradict it.
- Label every citation as verified-opened, suggested-but-not-opened, supplied-by-user, or model-memory. Never imply you opened a source unless a tool result confirms it.
- Return ‘insufficient evidence’ when the source does not support the claim. Do not fill the gap with a plausible guess.
What Makes Mistakes adds
The extension puts a visible reminder at the point where passive acceptance usually happens. Its Improve answer button starts the adversarial second pass in one click.
The remaining six steps are yours. A stronger second answer is still generated text; the fact-check begins when you inspect independent evidence.
Sources and further reading
Product details and guidance were checked against these first-party pages on July 29, 2026. Re-check current listings before making an install or high-stakes decision.
- Does ChatGPT tell the truth?
OpenAI Help Center
OpenAI's guidance on incorrect answers, fabricated citations, confidence, search tools, and verification.
- What are AI hallucinations?
Google Cloud
A plain-language definition of hallucinations, common causes, examples, and grounding strategies.
FAQ
Questions people ask
Can ChatGPT fact-check its own answer?
It can critique and revise its answer, especially when search tools or supplied evidence are available. That is useful, but it is not independent verification. Check the sources directly.
What sources should I trust when checking ChatGPT?
Prefer original and authoritative sources appropriate to the claim: legislation, regulators, standards bodies, official documentation, datasets, peer-reviewed papers, filings, and direct recordings or transcripts.
How do I know whether a ChatGPT citation is fake?
Search the exact metadata, open the original, and match the passage to the claim. A real-looking citation can be fabricated or point to a source that says something else.
Do I need to check every sentence?
No. Use consequence and novelty. Check the claims that are specific, current, surprising, cited, or capable of changing a decision.