This is a free AI hallucination checker you can use in any chatbot. Paste in an AI answer and the prompt breaks it into single claims, rates how confident it is in each, flags the ones you must verify yourself, and tells you what kind of source to check.
It won’t prove anything is true. It turns a wall of confident text into a short, ranked to-do list. Below: the full prompt, a second-opinion version for a different AI, a worked example, and the mistakes to avoid.
What the checker does (and what it can’t)
- It extracts every checkable claim: names, numbers, dates, quotes, citations, product features, rules.
- It rates each one High, Medium or Low confidence, with a reason.
- It forces the risky types into “must verify”: every quote, citation, statistic and anything that changes over time, no matter how sure the AI feels.
- It names the kind of source to check, like the official vendor page or the original study, instead of inventing a link.
Be clear about the limit: an AI checking AI is a filter, not proof. A model that invented a fact may happily “confirm” it, and a confidence rating is the model’s guess about itself, not a measurement. The checker’s real job is to point you at the few claims worth checking by hand. If you want to understand why the errors happen in the first place, read why AI makes things up.
The AI hallucination checker prompt
Copy the whole thing. Fill in the two bracketed sections, then send it.
You are a careful, skeptical fact-checker. Audit the AI answer below for made-up, outdated or unsupported facts. You did not write it and you are not defending it.
ORIGINAL QUESTION (for context):
[paste the question that produced the answer]
ANSWER TO CHECK:
[paste the full answer]
STEP 1: EXTRACT CLAIMS
List every checkable factual claim as a numbered list. Split long sentences so each item holds one claim. Include names, numbers, dates, statistics, quotes, citations, URLs, product features, laws or rules, and cause-and-effect statements. Skip opinions, advice and filler.
STEP 2: RATE EACH CLAIM
Make a table with these columns:
# | Claim (short) | Type | Confidence (High / Medium / Low) | Why | Must verify? (Yes / No) | Best source to check
Rating rules:
- High: widely documented, stable over time, and you would bet on it.
- Medium: probably right, but the exact figure, date or wording could be off.
- Low: very specific or obscure, you can't recall a reliable basis, or it may have changed recently.
- Every exact quote, citation, statistic, URL, and anything that changes over time (prices, plans, laws, versions, job titles) is "Must verify: Yes" no matter how confident you are.
- For "Best source to check," name the TYPE of source (official vendor page, the original study or its DOI, a government statistics site, the court record, the speaker's published remarks). Never invent a link.
STEP 3: RED FLAGS
Below the table, list any claim with a warning sign: a citation you can't confirm exists, a quote with no clear original, a suspiciously precise number, possibly outdated information, a specific-sounding detail with no clear source, or two claims that contradict each other.
STEP 4: CHECK THE QUESTION
Did the original question contain an assumption that might be false? Did the answer go along with it?
STEP 5: SUMMARY
- Overall risk: Low / Medium / High, with one sentence explaining why.
- The 3 claims I should verify first, in order of how much damage they'd do if wrong.
- For any claim you now believe is wrong, give a correction labeled "Correction (unverified)."
RULES
- Do not invent sources, links, page numbers or DOIs. If you don't know where something is documented, write "source unknown."
- "I'm not sure" is an acceptable rating. An honest messy report beats a clean-looking one.
- Do not rewrite or improve the answer. Only audit it.How to run the checker in four steps
- Paste both the question and the answer. The question lets the checker spot a false assumption the answer played along with.
- Use a fresh chat if you can. A new conversation isn’t anchored to the answer it just wrote. Running it in the same chat still works, but expect it to be kinder to itself.
- Read the “Must verify: Yes” rows first, starting with its top three.
- Check those rows yourself. Open the source type it suggested and confirm the claim. The 3-minute AI fact-check workflow walks through exactly how, including citations, numbers, quotes and dates.
The second-opinion version (for a different AI)
Different tools were trained differently, so they don’t always make the same mistake. Paste this into a second chatbot and compare. If you’re not sure which one to use, here’s how ChatGPT, Claude and Gemini compare.
I'm fact-checking an answer another AI gave me. Don't assume it's right.
Question: [paste the question]
Answer: [paste the answer]
1. List every specific claim (names, numbers, dates, quotes, citations, features).
2. For each one, say Agree, Disagree or Not sure, with one line of reasoning.
3. If you can search the web, link a page that supports or contradicts each claim. If you can't search, say so.
4. End with the claims where you disagree or aren't sure. Those are the ones I'll check myself.Two ways to make the second opinion more independent:
- Use a clean chat. In ChatGPT, a temporary chat set to Unpersonalized before your first message skips your memory and custom instructions, so it isn’t nudged by what it knows about you.
- Ask the same question more than once. Anthropic’s guide to reducing hallucinations suggests running the same prompt several times and comparing, because inconsistent answers can signal a hallucination. If a date changes between runs, check that date.
A worked example: checking a team briefing
Say you asked an AI for a short team briefing on a new workplace rule. The answer has some background, the date the rule takes effect, a statistic, a quote from the official who announced it, and a link to “the official guidance.” A useful checker run on that looks like this:
- The background sentences come back as High confidence, “Must verify: No.” Fine to skim.
- The effective date is rated Medium, marked must-verify because dates change, with “the agency’s official announcement” as the source to check.
- The statistic lands in Red Flags as suspiciously precise with no clear origin. Source to check: the agency’s own data page.
- The quote is must-verify by rule, with “the official’s published remarks or press release” as the source.
- The link is flagged because the checker can’t confirm that URL exists.
Your job is now three checks instead of forty sentences. Open the real guidance page, confirm the date and find the quote. If you can’t find the statistic in a minute, cut it.
When you have the source document: the strict version
If the answer was supposed to come from a document (a policy, a report, meeting notes), skip confidence ratings. Anthropic recommends having the model find a supporting quote for each claim and drop any claim it can’t support:
Below are a SOURCE DOCUMENT and an ANSWER that was supposed to be based on it.
For each factual claim in the answer:
1. Quote the exact sentence from the source that supports it.
2. If no sentence supports it, label the claim UNSUPPORTED.
3. If the source says something different, label it CONTRADICTED and quote the source.
Use only the source document. Do not use outside knowledge.
SOURCE DOCUMENT:
[paste]
ANSWER:
[paste]This is much closer to real verification: every “supported” claim comes with a quote you can find with Ctrl+F.
Mistakes that make the checker useless
- Treating “High confidence” as verified. It means the model feels sure. Plenty of hallucinations are delivered with total confidence.
- Checking only the red flags. Anything you’ll act on, publish or send to your boss deserves a look, flagged or not.
- Accepting the “corrections.” A correction from the same model is another unverified claim. That’s why the prompt labels them.
- Dropping the original question. Without it, the checker can’t catch an answer that went along with a false premise.
If you’d rather prevent errors than catch them, pair this with the 7 AI accuracy fixes. ChatGPT users can also bake the same rules into their account with these accuracy settings.
Try it now
Find the last AI answer you used for something real: an email, a report, a decision. Paste it and its question into the checker, then check just one of the top three claims it flags. That single check will tell you more about how far to trust your AI than any amount of reading.
🎁 Free download: The AI Fact-Check Kit
The 3 prompts that catch fabricated facts before they cost you, plus the verification workflow for research that has to be right — in one free Notion template. Free with your email address.
Related guides
- Why AI Makes Things Up: Hallucinations Explained
- The AI Fact-Check Workflow: Verify Any Answer in 3 Minutes
- Your AI Is Lying to You: 3 Prompts to Catch Hallucinations
- How to Improve AI Accuracy: 7 Fixes That Actually Work
- The AI Tool I Use Most (And the 5 Others I Use Too)
Frequently asked questions
Can one AI reliably check another AI's answer?
Not reliably enough to count as proof. A second model can catch errors the first one made, and disagreement between tools is a useful warning sign, but both can be wrong in the same way. Use the checker to find what to verify, then verify it against a real source.
Does the hallucination checker prompt work in ChatGPT, Claude and Gemini?
Yes. It's plain text, so it works in any chat tool, including free plans. It works better when the tool can search the web, because it can then link pages that support or contradict each claim.
Should I run the checker in the same chat as the original answer?
A fresh chat or a different tool is better, because it isn't anchored to what it just wrote. Running it in the same chat still helps, but expect it to go easier on its own work.
