Lietuviškai
Deutsch
Français
Русский
Español
Polski
Italiano
Українська
Čeština
Português
日本語
Latviešu valoda
Eesti
Magyar
Bahasa Indonesia
Türkçe
हिन्दी
Ελληνικά
العربية
ไทย
Tiếng Việt
Română
Bahasa Melayu
বাংলা
AI fact-checking is the practice of using an assistant to test a claim against sources before you repeat it: in a post, a report, a client e-mail or a conversation at dinner. Used well, it is fast and surprisingly rigorous, because the assistant can search, read the original document and tell you exactly which sentence supports or contradicts the claim. Used badly, it is worse than no check at all, because a model asked “is this true?” without sources will answer from memory in a confident tone, and a confident wrong answer is the thing you were trying to avoid.
This guide shows a method that works: turn the claim into checkable parts, make the assistant search and cite, read the sources yourself, and know which kinds of claims an assistant cannot settle. It uses Ask Mio’s Research mode in the examples.
AI fact-checking: why “is this true?” is the wrong question
A language model answers from two places: what it absorbed during training and what you put in front of it. Ask it whether a statistic is correct with no sources attached, and it answers from training, which has a cutoff date, contains errors and is reproduced from memory. The answer sounds the same whether it is right or wrong. That is the mechanism behind hallucinations, and if you have not read about it, start with what AI hallucinations are and how to catch them.
AI fact-checking works when the assistant is forced to answer from the second place: documents and pages it has actually retrieved for this question, which it then cites so you can read them. Research mode does the searching and shows the sources. Your job is to ask in a way that makes the sources carry the answer, and then to read them.
Step 1: Break the claim into checkable parts
Most claims that spread are bundles. “The new EU rule bans X from 2027 and will cost small businesses €2 billion” contains at least four separate assertions: that there is a rule, that it bans X, that it applies from 2027, and that a cost estimate of €2 billion exists and refers to small businesses. Each has a different kind of source and a different chance of being wrong.
Ask the assistant to do the splitting:
Break this claim into its separate factual assertions, one per line, and for each one say what kind of source would settle it (official text, statistics office, peer-reviewed study, company statement, news report). Do not evaluate anything yet.
This step alone catches a lot. Claims that cannot be split into checkable assertions, such as “experts say”, “it is well known that” or “studies show”, are not facts yet; they are invitations to find out who said it.
Step 2: Make the assistant search and cite
Now check each assertion with a prompt that demands sources:
Check this assertion using web search. Find the primary source if one exists (the official text, the original study, the statistics release), not a news article about it. Quote the exact sentence from the source that supports or contradicts the assertion, give the link, and say whether the source is primary or secondary. If you cannot find a source, say so; do not answer from memory.
Three details in that prompt do the work. “Primary source” steers away from the chain of articles quoting each other. “Quote the exact sentence” makes it impossible to paraphrase a source into saying something it did not. “If you cannot find a source, say so” gives the assistant permission to return empty-handed, which is the honest result for many claims.
Our guide to researching with AI and getting cited sources covers the mechanics of Research mode; the prompt above is the fact-checking variant.
Step 3: Read the sources yourself
This is the step people skip, and it is the one that matters. Open the link. Find the quoted sentence. Then read around it, because the three most common ways a true quotation supports a false claim are all about context:
- Scope. The figure is real but refers to a different population, year or country than the claim implies.
- Tense. A proposal is reported as a decision; a draft as a law; a pilot as a rollout.
- Attribution. The source is a press release, an opinion column or a lobby group’s estimate, presented as if it were an official measurement.
If the page is long, upload it or paste the URL and ask: “Where in this document is the figure X, and what population and period does it refer to?” The assistant becomes a reading aid for a source you have in front of you, which is its most reliable role. If the page needs a real browser to render, the live browser tool fetches it; the difference between that and a plain page read is explained in AI live browser vs web page reader.
Step 4: Weigh the source
Not every source settles a claim. A quick hierarchy:
| Source type | Settles the claim? | Watch for |
|---|---|---|
| Official legal text, regulation, court ruling | Yes, for what the text says | Version and date; proposals vs adopted text |
| National or EU statistics office release | Yes, for the figure as defined | Definitions, revisions, reference period |
| Peer-reviewed study | Strong evidence, not proof | Sample size, what was actually measured, later studies |
| Company or organisation statement | Yes, for what they said; not for whether it is true | Marketing framing, missing numbers |
| Reputable news report | Usually reliable for events; secondary for figures | Which primary source it relies on |
| Fact-checking organisation | Useful summary with sources | Read their sources, not just the verdict |
| Social media post, forum, anonymous blog | No | Treat as a claim to check, not a source |
| The assistant’s own memory | No | Confident tone, no link, possibly outdated |
Professional fact-checkers work under published standards; the International Fact-Checking Network’s code of principles is a short read and a good model for your own habits: non-partisanship, source transparency, and corrections when wrong. In the EU, the Digital Services Act now obliges large platforms to assess and reduce the spread of disinformation, which is one reason platform labels and context notes have become more common; they are useful pointers, but they are not a substitute for reading the source.
Claims an assistant cannot settle
Be clear about the limits, because they are where confident errors live.
Very recent events
Models have a training cutoff, and even with web search, the first hours of a developing story are full of contradictory reports. An assistant can collect what has been published; it cannot tell you which early report will hold. Our article on AI knowledge cutoffs explains why “the model said so” is not evidence for anything recent.
Images, video and audio
An assistant can describe an image and flag obvious inconsistencies, but it cannot reliably tell a manipulated image from a real one, and it cannot verify that a video shows what the caption says. Reverse image search, metadata and provenance tools are the right instruments; see our guide to spotting deepfake voices and videos.
Private or paywalled information
If the source is behind a login, inside a company or in a print archive, the assistant cannot retrieve it. It may offer a secondary source instead; note the downgrade.
Predictions and opinions
“This policy will cost jobs” is not a fact to check; it is a forecast. What can be checked is who made it, on what model, and what comparable forecasts came true. Ask for that instead.
Claims about what “most experts” think
Checkable only through surveys of experts or position statements by professional bodies. Ask for those specifically; otherwise the assistant will summarise a vibe.
A worked example
Claim seen in a newsletter: “Remote workers are 13% more productive, according to a Stanford study.” Split: a study exists; it is from Stanford; it found 13%; the finding is about remote workers in general.
Search with citations finds the primary source: a published economics paper on a call-centre experiment in China, with a 13% performance increase for employees working from home in that specific setting. So: the study exists, the affiliation is right, the figure is right, and the generalisation is wrong. The study measured one company, one job type, one country, a decade ago, and later work by the same authors reported different effects for different arrangements. The honest version of the claim is narrower, and the assistant could only produce it because it was asked to quote the source’s own description of its sample.
That pattern, true numbers attached to an overstretched claim, is the most common thing AI fact-checking catches.
Prompts for everyday situations
Before you share a post
“Here is a post I am about to share. List its factual claims, check each with web search and primary sources, and tell me which ones hold, which are partly true, and which you could not verify.”
Checking your own draft
“Here is my article. Mark every sentence that contains a checkable fact, and for each one tell me whether I cited a source, and whether the source actually supports the sentence as written.” This is a strong editing pass for anything you publish.
Checking a supplier or competitor claim
“This company claims its product is certified to standard X. Find the certification body’s register and tell me whether the company or product appears in it, with a link.”
Checking a statistic in a meeting
“Someone said unemployment in country Y is at a 20-year low. Find the latest release from the national statistics office, quote the headline figure and the period, and say whether the 20-year claim is supported by the series.”
In Ask Mio, a research answer with search costs a few points; Research mode with web search is included from the Chat plan at €5 a month plus VAT. Keep a “fact checks” project with your preferred sources and the prompts above, and every check starts from the same standard.
Habits that keep the method honest
- Never accept a verdict without a link you opened.
- Prefer the primary source even when a summary is more convenient.
- Record the date you checked; figures and rules change.
- When the assistant says it cannot find a source, treat that as a result, not a failure to try harder with leading prompts.
- Correct yourself publicly when you were wrong; it is the one habit that makes the rest credible.
Frequently Asked Questions
Can AI fact-check reliably?
It can, when it is made to search, retrieve and quote sources rather than answer from memory, and when you read the sources yourself. Asked “is this true?” without sources, an assistant answers from training data in a confident tone, which is not a check. The method, not the model, decides whether AI fact-checking is reliable.
Which sources should an AI fact-check rely on?
Primary sources first: official legal texts, statistics office releases, the original study, the organisation’s own statement. Reputable news is reliable for events but secondary for figures. Social media posts and the assistant’s own memory are claims, not sources. Ask explicitly for the primary source and for the exact sentence that supports or contradicts the claim.
Why did the assistant give a confident answer that turned out wrong?
Because it answered from training data rather than from a retrieved source, or because it paraphrased a source into saying more than it did. Both are prevented by the same prompt rules: search, quote the exact sentence, give the link, and say “not found” when nothing supports the claim. Then open the link and read around the quote.
Can AI detect fake images or deepfakes?
Not reliably. An assistant can describe an image and point out visible inconsistencies, but it cannot certify that an image or video is genuine. Use reverse image search, provenance metadata and specialised tools, and treat any image-based claim as unverified until the original context is found.
How do I fact-check something that happened today?
With caution. Web search finds the early reports, but early reports contradict each other and the model’s own knowledge has a cutoff. Collect what the main outlets and official channels have said, note the time, and avoid stating a figure or a cause until a primary source, such as the authority involved, has published one.
Which Ask Mio plan do I need for fact-checking?
Research mode with web search and cited sources is included from the Chat plan at €5 a month plus VAT, which also adds file uploads for checking long documents. The free plan covers Chat and Write without web search, so it is fine for splitting claims and editing but not for retrieving sources. Details are on the live pricing page.
The Bottom Line
AI fact-checking is a method: split the claim, make the assistant search and quote primary sources, read those sources yourself, weigh what kind of source each is, and recognise the claims no assistant can settle. Follow it and an assistant becomes a fast, careful research partner. Skip the reading step and it becomes a confident voice that is sometimes wrong, which is exactly what you were checking against. Research mode with cited sources starts with the Chat plan; compare the Ask Mio plans or try the free plan for the claim-splitting steps.