Check a claim you saw online, with AI as a helper, not the judge

Use it when you saw a post, screenshot or headline, such as a health tip, a statistic, a quote or a piece of breaking news, and you want to know whether to believe it, share it or act on it.

Not when you are checking an answer an AI gave you. For that, use Check before you trust: a five-minute routine. For a health tip that would change what you take, eat or stop doing, ask a pharmacist or doctor as well.

You need the post itself or a screenshot, an assistant, a search engine and about ten minutes. The steps and labels are our suggestion.

Why the assistant is a helper, not the judge

Answers about news are often wrong. In a 2025 study by public broadcasters, 45% of AI answers about news had at least one significant issue (EBU/BBC, 2025, field study).
Citations are often wrong, and confidently so. Eight AI search tools gave incorrect answers to more than 60% of queries asking which news article an excerpt came from (Tow Center, 2025, field study). More on both studies: AI answers or search results: when to use which.
A verdict can make you worse at judging. In an experiment planned and registered in advance, an AI model's fact checks spotted 90% of false headlines but did not improve people's ability to tell true from false. Where it mislabelled a true headline as false, belief in that headline fell, and where it was unsure about a false one, belief rose. Fact checks written by people did improve discernment (DeVerna et al., 2024, peer-reviewed, PNAS).
A conversation can help. In two experiments with 2,190 Americans, a dialogue with GPT-4 Turbo (an AI model from OpenAI, the maker of ChatGPT) about a conspiracy theory each person believed, described in their own words, cut that belief by about 20%, and the effect lasted at least two months. A professional fact checker rated 127 of the model's 128 claims true and none false (Costello, Pennycook and Rand, 2024, peer-reviewed, Science).

So an assistant is good at breaking a claim down, explaining it and suggesting where to look. Its verdict on a specific, recent claim is the part to distrust.

How professional fact checkers do it

Researchers at Stanford watched 10 professional fact checkers, 10 historians with PhDs and 25 Stanford undergraduates judge real websites. Historians and students read "vertically", staying on a site to judge it, and often fell for official-looking logos and domain names (web addresses). Fact checkers "read laterally, leaving a site after a quick scan and opening up new browser tabs in order to judge the credibility of the original site", and reached better conclusions "in a fraction of the time" (Wineburg and McGrew, 2017, working paper, later peer-reviewed). The free Civic Online Reasoning lessons, now from the Digital Inquiry Group, teach this, starting with "Intro to Lateral Reading".

Steps

#	Step	Who	Time	Check
1	Pause. Write the claim down exactly: who said it, what it says, any number, when, where, and where you saw it. Save the link or screenshot.	Person	2 min	One or two sentences you could hand to someone else.
2	Paste the claim into the prompt below. Ask for checkable parts and the sources that would settle them, not a verdict.	Person and AI	2 min	Each part has a kind of source next to it. If it fails: if the assistant rules anyway, ignore the verdict and keep the list.
3	Read laterally. Leave the post and search the account, site or outlet by name. Read what others say about who runs it and who funds it.	Person	3 min	You can say in one line who is behind the claim.
4	Find the original: the study, the official figure, the full quote or the full video. Check that it says what the post says, with the same number, date and scope.	Person	3 min	You opened the original and found the line. If it fails: if you cannot find it, the claim is unverified, however often it was shared.
5	Search for an existing fact check in Google Fact Check Explorer and on a fact-checking site you trust.	Person	1 min	A fact check found, or "none found".
6	If there is a picture or video, run a reverse image search (search with the picture itself to find other copies): Was this image made or changed by AI? A five-minute check.	Person	2 min	The earliest copy you found, or "none".
7	Label it: supported by the original, contradicted, missing context, or unverified. Share only the first, and with the original's link.	Person	1 min	One label and a link to what you read.

Google says you can search a topic in Fact Check Explorer to find fact checks by independent organisations, and upload or paste an image link to see whether it appears in one (Google, April 2024).

Claim breakdown prompt · v1.0

Here is a claim I saw online. Do not tell me whether it is true.

Claim, exactly as written: <CLAIM_TEXT>
Where I saw it: <WHERE_SEEN>

1. Split it into separate checkable parts: who, what, how much, when, where.
2. For each part, say what kind of source would settle it (for example the
   original study, an official statistic, the full speech or video) and who
   would publish that source.
3. Give search terms I can use to find each source myself.
4. List what is vague or missing: no date, no named source, a number with no
   base, a quote with no context.
Name kinds of sources, not specific articles, unless you give a link I can open.

Replace the following:

<CLAIM_TEXT>: the claim from step 1, word for word, including any number.
<WHERE_SEEN>: such as a screenshot shared in a family group chat, no link.

Breaking news. In the first hours after a big event, wait. A long-standing handbook for news readers opens with "In the immediate aftermath, news outlets will get it wrong", and warns that "Big news brings out the fakers" (On the Media, 2013). Assistant answers can be out of date too. Check again the next day, and prefer outlets close to the event.

Worked example (illustrative). A screenshot says "New university study: 40% of office workers now use AI every day." The assistant splits it into who (which university, and who ran it?), what ("use AI" how?), how much (40% of whom?) and when ("now" is which year?). Searching the account shows it is a marketing newsletter. The original turns out to be a vendor survey of one country that found 40% use AI "at least weekly", with no university involved. Label: missing context. Do not share the screenshot.

How to tell it worked

You can state the claim in one or two exact sentences.
You opened the original source, or you labelled the claim unverified.
You know who is behind the account or site that posted it.
Your label rests on what you read, not on the assistant's opinion.
It has failed if the only evidence you have is the assistant saying it is true or false.

Sources

Wineburg S and McGrew S. Lateral reading: reading less and learning more when evaluating digital information. Stanford History Education Group working paper, 2017; later published in Teachers College Record. Study. Link
Digital Inquiry Group. Civic Online Reasoning. Free lessons. Link
EBU and BBC. News Integrity in AI Assistants, 2025. Field study. Link
Jaźwińska K and Chandrasekar A. AI search has a citation problem. Tow Center, Columbia Journalism Review, March 2025. Field study. Link
DeVerna MR et al. Fact-checking information from large language models can decrease headline discernment. PNAS, 2024. Peer-reviewed experiment. Link
Costello TH, Pennycook G and Rand DG. Durably reducing conspiracy beliefs through dialogues with AI. Science, 2024. Peer-reviewed experiment. Link
Google. 4 ways to use Search to check facts, images and sources online, April 2024. Vendor page. Link
On the Media (WNYC). The Breaking News Consumer's Handbook, 2013. Media guide. Link

Changes

v1.010 Oct 2026first version

Why the assistant is a helper, not the judge

How professional fact checkers do it

Steps

How to tell it worked

Sources

Changes

Related