The source exists; the number is rounded or a year out
Fix the detail. This is the most common result and the reason to check numbers you will repeat.
How to fact-check ChatGPT
How to fact-check ChatGPT in five minutes, claim by claim.
ChatGPT writes paragraphs; facts live in sentences. Fact-checking means breaking the answer into its claims, deciding which ones matter, and finding a source you can open for each of those, ideally with a model that was not the author doing the finding. This page is the procedure, the prompts, and the tools inside MultipleChat that do most of the work: Perplexity for sources, a second model for judgement, and Auto Verification on paid plans.
Free to start. Auto Verification automates the source check on paid plans.
The procedure
Five steps, claim by claim
Split the answer into claims
Paste it and ask: "List every factual claim in this as a numbered list, one sentence each." Opinions and explanations are not claims; numbers, names, dates, quotes and "studies show" are.
Decide which claims matter
The ones you will repeat, publish, or act on. Mark them. The rest can stay unchecked; checking everything is how people give up.
Find a source for each marked claim, with a model that did not write it
Send the list to Perplexity: "For each claim, find a source I can open and quote the relevant sentence. Say 'no source found' where there is none." Retrieval, not memory.
Open the sources that matter
Read the sentence the claim rests on. A source that exists but says something different is the most common failure, and only you will catch it.
Record the verdicts
Supported, unsupported, contradicted, with the link, in the Project. Now you can defend the answer, and the next time the topic comes up the work is done.
Which model for which step
Author, finder, judge
| Step | Model | Why |
|---|---|---|
| Splitting into claims | Any | Structural task; every model does it well. |
| Finding sources | Perplexity (Sonar; Sonar Pro on paid plans) | Search-native, cites pages you can open. |
| Judging support | Claude or ChatGPT, a different one from the author | Reads the claim against the quoted source; a different model does not defend the original. |
| Automated pass | Auto Verification (paid plans) | An independent verification of factual claims when confidence matters, without orchestrating the steps yourself. |
| Final read | You | Opening the source is the step no model replaces. |
Do not ask the author
Asking ChatGPT "are you sure?" gets you reassurance. Asking a different model with the source in hand gets you a verdict.
Common outcomes
What the check usually finds
A real study, misread
The dangerous one. Only opening the source catches it.
A plausible specific with no origin
Cut it, or rewrite as an estimate marked as such.
The source says the opposite
Rarer than people expect; usually recency. Check the date.
Free is the start. The latest models are the difference.
Checking answers is where the latest models earn their price.
The free plan gives you the major model families and enough messages to make second opinions a habit, which is the whole point of this page. The paid plans add the latest model in each family, which disagree in more useful ways and invent less; Perplexity Sonar Pro for sourced facts; and Auto Verification, an independent verification pass that runs when factual confidence matters, without you orchestrating it.
Newer models hold your whole Project in view instead of the last few messages, reason through trade-offs instead of picking the first plausible answer, and are wrong less often and with less confidence. For the questions on this page, that is the difference between advice that sounds right and advice you can act on.
Pro from $20 a month, cancel any time. The free plan stays free.
- Your first Project and file uploads
- The major AI families to try the workflow
- A small daily message allowance
- The latest ChatGPT, Claude, Gemini, Grok and Perplexity models, all on the same Project
- Far more messages a day, so a working session does not stop halfway
- Modes where several models draft, challenge and verify each other's answers
- Auto Verification: an independent check on factual claims
- Perplexity Sonar Pro for sourced, current facts
- Collaboration modes where models draft, challenge and verify each other
Start here
The fact-checking prompts
In order.
SplitList every factual claim in this answer as a numbered list, one sentence each. Leave out opinions and explanations.
PrioritiseWhich of these claims would matter most if wrong, given I will use this for [purpose]?
Find (Perplexity)For each numbered claim, find a source I can open and quote the sentence that supports it. Write "no source found" where there is none.
Judge (other model)Here are claims with quoted sources. For each, say supported, unsupported or contradicted, and why.
RecordProduce a table: claim, verdict, source link, note. I will keep this in the Project.
RewriteRewrite the original answer keeping only supported claims, marking estimates as estimates.
Honest answers
What people ask
Can ChatGPT fact-check itself?
Poorly. It tends to reaffirm its own answer. Use a different model with a retrieved source in hand.
Is Perplexity reliable for fact-checking?
It is reliable at finding and citing pages. Whether the page says what the claim says is for you, or a second model, to judge. Open the source.
How long does this take?
Five to ten minutes for a typical answer with five marked claims. Auto Verification on paid plans shortens the finding step.
What about claims with no possible source?
Predictions, opinions and private facts cannot be fact-checked. Label them as such and do not present them as facts.
Can I fact-check a whole document?
Yes. Upload it to a Project and run the same procedure section by section. See analysing files with AI.
What is free?
Perplexity Sonar, the major model families and a daily message allowance. Sonar Pro, Auto Verification and the latest models are on paid plans.
Split it. Source it. Open it. Record it.
Five minutes in MultipleChat, and the answer becomes something you can defend. Free to start.
Also: does AI make things up? · is ChatGPT always right? · market research with AI