How to fact-check ChatGPT answers: a copyable source-check prompt (tested on 2 AIs)
An AI answer reads well, so it feels true. But a confident sentence isn't a checked one. Here is a fact-check you can do in about two minutes: paste the answer, paste a source you trust, and make the AI prove every claim with a quote from that source. We tested it on two models, with one mistake planted on purpose.
Why ChatGPT sounds sure when it's wrong
In September 2025, three OpenAI researchers and one from Georgia Tech published Why Language Models Hallucinate. Their explanation is simple: models are trained and graded in ways that reward a confident guess more than an honest "I don't know".
Language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty.
Kalai, Nachum, Vempala and Zhang, arXiv, September 4, 2025
So asking the same chat "are you sure?" is not a real check: it has nothing new to compare against. A real check needs a source, and proof you can see.
Our test: one letter, one summary, three claims
We wrote a fictional letter from a school about a Grade 4 field trip, then an AI-style summary of it. We planted three kinds of claims: one right, one with the wrong date, and one the letter never says. The letter and the summary are fictional. The answers below are real.
| Claim in the summary | What the letter says | Expected label |
|---|---|---|
| The trip is on Friday, October 30 | Thursday, October 29 | WRONG |
| The signed slip and $12 are due Monday, October 26 | The same, word for word | VERIFIED |
| Parents are welcome as volunteer chaperones | Nothing about parents | CAN'T VERIFY |
The prompt (copy it)
Fact-check the ANSWER below using only the SOURCE below. 1. Split the ANSWER into separate claims, one fact per line. 2. Give each claim one label: VERIFIED = the SOURCE says the same thing. WRONG = the SOURCE says something different. CAN'T VERIFY = the SOURCE doesn't say. 3. For each label, quote the exact words from the SOURCE. No quote, no VERIFIED. 4. For each WRONG claim, write the corrected version, using only the SOURCE. Do not use outside knowledge. Do not add new facts. Answer as a table: Claim | Label | Exact quote from the SOURCE | Fix ANSWER: [PASTE THE AI ANSWER] SOURCE: [PASTE THE PAGE, EMAIL OR DOCUMENT YOU TRUST]
- Paste the answer
Copy the AI's answer under ANSWER, as it is.
- Paste the source
Under SOURCE, paste the page, email or document you trust. Not another AI answer.
- Read the labels
WRONG comes with a fix. CAN'T VERIFY means your source doesn't say. VERIFIED comes with a quote.
- Find the quotes yourself
Search your source for each quote (Ctrl+F or Cmd+F). If you can't find it, it isn't verified.
The real answers (October 18, 2026)
We sent the exact prompt, with the letter and the summary, to two models in a fresh session, with no web access and no memory.
Claude Opus 5.5 (fresh chat, no tools)
| Claim | Label | Exact quote from the SOURCE | Fix | |---|---|---|---| | The trip is for Grade 4. | VERIFIED | "Our Grade 4 classes will visit the science centre" | — | | The trip is to the science centre. | VERIFIED | "Our Grade 4 classes will visit the science centre" | — | | The trip is on Friday, October 30. | WRONG | "will visit the science centre on Thursday, October 29" | The trip is on Thursday, October 29. | | The signed permission slip is due Monday, October 26. | VERIFIED | "Please send back the signed permission slip and $12 by Monday, October 26." | — | | $12 is due Monday, October 26. | VERIFIED | "Please send back the signed permission slip and $12 by Monday, October 26." | — | | Parents are welcome to come along as volunteer chaperones. | CAN'T VERIFY | None. The SOURCE doesn't mention parents coming along, volunteers or chaperones. | — |
GPT-6.1 via OpenAI Codex (not the ChatGPT app)
| Claim | Label | Exact quote from the SOURCE | Fix | |---|---|---|---| | The Grade 4 classes will visit the science centre. | VERIFIED | “Our Grade 4 classes will visit the science centre on Thursday, October 29.” | — | | The trip is on Friday, October 30. | WRONG | “Our Grade 4 classes will visit the science centre on Thursday, October 29.” | The trip is on Thursday, October 29. | | The signed permission slip is due Monday, October 26. | VERIFIED | “Please send back the signed permission slip and $12 by Monday, October 26.” | — | | The $12 payment is due Monday, October 26. | VERIFIED | “Please send back the signed permission slip and $12 by Monday, October 26.” | — | | Parents are welcome to come along as volunteer chaperones. | CAN'T VERIFY | No quote available: the SOURCE does not mention parent chaperones. | — |
Both models caught all three: the wrong date, with the right one quoted; the deadline, verified word for word; and the chaperone line, marked CAN'T VERIFY. Claude split the summary into six claims, GPT-6.1 into five. Same verdicts.
The source check, layer by layer — AI answer, Your trusted source, One label per claim, An exact quote, You open the source
No document to paste? Get the source first
- Ask for sources
Ask: "For each claim, give me the source, with a link."
- Open every link
A link can be real and still not say what the answer claims.
- Copy the paragraph
Copy the part of the page that matters.
- Run the same prompt
Paste that paragraph as the SOURCE. Now you're checking the page, not the AI's memory.
When it's worth two minutes
Not for a dinner idea. Do it when an answer goes somewhere: a date in your calendar, an email you'll send, a number you'll repeat, anything you'll act on. And read CAN'T VERIFY as "not yet": the claim may be true, but your source doesn't say it, so check before you use it.
More tested prompts, in English and in French: the Lucarne prompt library.
Sources
Frequently asked questions
How do I fact-check a ChatGPT answer?
Paste the answer and a source you trust into the source-check prompt. It splits the answer into claims and labels each one VERIFIED, WRONG or CAN'T VERIFY, with an exact quote from your source. Then search your source for each quote yourself. In our test on October 18, 2026, it caught a planted wrong date on both Claude and GPT-6.1.
Can I get ChatGPT to fact-check itself?
Asking "are you sure?" in the same chat is not a real check: the model has nothing new to compare against. Give it a source (the page, email or document) and require a quote for every VERIFIED claim. With no document, ask for a link per claim, then open each link yourself.
Why does ChatGPT make things up?
A September 2025 paper by OpenAI researchers argues that training and evaluation reward guessing over admitting uncertainty, so models learn to answer confidently even when they don't know. A confident tone is not proof.
Go further
- Prompt: fact-check a ChatGPT answer against a source you trustPaste an AI answer and a source you trust. The AI splits the answer into claims and labels each VERIFIED, WRONG or CAN'T VERIFY, with an exact quote. Tested October 18, 2026 on Claude and GPT-6.1: both caught a planted wrong date.See the prompt
- Prompt: make the AI check its own draft against your rulesPaste a draft and your rules; the AI lists every rule the draft breaks, then rewrites it. Tested on a generic template: it flagged the invented content, the missing facts and all 14 placeholders.See the prompt
- How to use ChatGPT effectively: the 5-part briefing prompt (real before and after)A one-line request got a 346-word template with 14 blanks. The same request with a five-part brief got an email ready to send. The briefing prompt to copy, the line that makes AI ask before it guesses, and how to save your brief.Read the article
Where should you start with AI?
The free assessment gives you, in 4 minutes, the 3 tasks to automate first in your business.