How we decide

Testing a post, article or claim

We don't look up what someone else concluded. We investigate each claim ourselves, in four steps, using AI models for the research. AI can make mistakes, so each step is designed to catch the errors of the others.

1. Separate the claims. We split the piece into neutral, standalone statements of one fact each and remove who said them, so a speaker's reputation can't sway the research. Every item in a list, every name, number and date, and anything the piece takes for granted gets tested, so nothing is skipped for looking minor or unlikely. A second reader, who sees only the text and the claims, sends the list back if a claim is softer than what was said, adds something, or leaves something out.
2. Try to prove it, and try to disprove it. For each claim, two separate researchers work with searches of their own, without seeing each other's findings. One looks for the strongest evidence in favour. The other tries to break the claim: it finds the original source, checks numbers and dates, looks for footage that's been recycled, and searches in the local language and among outlets with different stakes.
3. Weigh both files. A judge that did no research reads both and decides. It grades every source, reviews who each one speaks for rather than taking the researchers' word for it, and can only set evidence aside by giving a reason, so it can't quietly ignore what the challenger found.
4. Apply the evidence rules. Finally, the rules below decide whether the result is strong enough to stand.

• Confirmed

  Every part was tested. Independent sources or a primary record support it, and our attempt to disprove it found nothing that outweighs them.

• Refuted

  Whoever makes a claim has to prove it. At least one part is contradicted by a primary record or independent sources, or comes with no proof and our research found none. One false part makes the whole of it false.

Whoever makes a claim has to prove it. We look for that proof as hard as we look for the disproof, but if neither the piece nor our research produces it, the claim hasn't been proved, and it's Refuted. We say plainly which kind it is: “Contradicted by the record” and “No proof given, and we found none” are different findings.

A claim is never rated false because our own research failed. We try again, and if we still can't check it, we give no result rather than a guess.

Until November 2026 we also had a third verdict, Unconfirmed, for claims the evidence didn't settle. Results from then still show it. Checking them again gives Confirmed or Refuted.

Opinions, jokes, predictions and personal statements get no rating, because they can't be true or false.

A verdict on a claim isn't a verdict on the poster

People post claims for all sorts of reasons, including to say they're false. So we work out whether a post asserts a claim, passes it on, debunks it, or leaves that unclear, from its wording, its quotation marks and the author's bio. The researchers never see any of this, or who posted it.

Whenever a post isn't clearly asserting a claim, we say so next to the verdict, on every plan. A “Refuted” on a post that's debunking a rumour means the rumour is false, not the post. News outlets are treated a little differently: reporting what someone said counts as asserting that they said it.

The evidence rules

These are enforced in code, not left to the AI's judgement:

• Nothing is confirmed without an attempt to disprove it. If the attempt to disprove a claim didn't run, we don't rate the claim.
• Every quote must be on the page. We check that the exact words appear on the page they're cited from, and remove the citation if they don't. The one exception is PDFs, whose text we can't read yet. They're marked as unverified in exported dossiers.
• We count origins, not articles. Ten outlets repeating one wire story or one official statement count as one source, and so does a report that relies on another.
• A report about a record isn't the record. An article citing data, a law or an official document can't count, and we won't repeat its figures, unless the original is opened. For trade between countries, we read what each country itself filed with the UN.
• Evidence has to match the claim. Imports aren't exports, a law on foreign firms says nothing about national ones, and smuggled goods aren't approved trade.
• A claim of change needs a before. “Imports were cut” takes for granted that there were imports to cut, so we test that as well, in words that don't assume the cut happened. A change can't be Confirmed unless how things stood before is too.
• Only direct evidence counts. A source has to speak to the claim as written. One that's merely consistent with it, or doesn't mention it, stays on the record but can't decide the verdict. Anything the judge didn't grade doesn't count either.
• The bar for proof. “Confirmed” needs direct evidence from a primary record, such as an original document or dataset, or from two independent origins. A claim that falls short of it hasn't been proved, so it's Refuted, and we say it wasn't proved rather than shown false.
• Two separate records settle it. If two separate primary records speak directly to a claim, and nothing direct points the other way, that decides the verdict, and we tell you when it does. That can leave out periods the records don't cover.
• Fact-checks are leads, not proof. We read them to find evidence, but another organisation's verdict never counts towards ours. Search results are leads too: we go by the page itself, never by what a search shows of it.
• Nobody can vouch for themselves. We go by where a statement comes from, not who carries it: a news article quoting a government official is that government's statement. A government, party, company or armed group backing its own claim doesn't count as corroboration, though an admission against its own interest does. What a country files with the UN about its trade is routine statistics, so it counts as a record. On X, only posts from official accounts and news organisations we've checked by hand can count, each as one report from that source. When a claim is about dealings between two states, we search both states' official accounts ourselves, so a researcher always has what they said to work from.
• Old acts can't settle a claim about now. If a claim says how things stand today, or that something has been under way lately, a source more than a year old doesn't count, whichever way it points and even if it's a state's own word. We keep it on the record and say why.
• How two states get on is a trend, not one statement. On a claim about two states, a separate researcher, who isn't told the claim, goes through what's happened between them: ambassadors, expulsions, detainees, protests, trade and visa measures, and steps towards reconciliation. You see that timeline with the verdict. A formal step by one against the other, such as a recall, an expulsion, a formal protest or a measure aimed at it, that a source from the last year shows is still in place speaks to how they stand, even if no source calls relations tense. A government saying relations are fine doesn't settle it. If nothing formal is open and they're dealing normally, a claim of tension fails. A call or a visit doesn't close a dispute: only the thing itself ending does, such as an ambassador returning or a detainee being freed.
• A source with a stake doesn't count on its own. On a claim about a state, that means a source run by that state, based in its rival, or that we've listed as hostile to it. We keep it on the record and say why, and it counts if it cuts against its own side or the other side says the same thing. This runs both ways between rivals. We keep a list of outlets, who owns them and their editorial line, and outlets with one owner count as one source.
• Unusual claims need more. A claim that would break sharply with how things normally stand, such as close neighbours cutting ties, needs three independent origins rather than two, or a primary record. If something like that would be public and nobody without a stake reports it, we rate it false.
• A state can't hide a measure by not announcing it. A measure applied through verbal instructions leaves no official text, so two separate reports from sources with no stake can confirm it, even if the state denies it. We tell you when a verdict rests on that.
• A state's word settles what it decided, and nothing else. For a decision a state took, such as cutting ties or passing a law, or how its relations with another state stand, we go to its ministries, official press agency, spokesperson or gazette, and that statement settles it. What a state claims happened on the ground is still only a claim to test.
• A denial proves nothing by itself. When the party concerned denies something, that can't make a claim false, and it can't hold back one that independent evidence establishes. A state's standing record of its own acts is different: its embassy's page isn't a denial, so it can show that claim about the embassy is false.
• Some things can't happen quietly. Two states can't cut ties without anyone announcing it. If we really look and find no announcement or report of an act like that, we rate the claim false, say why, and look again soon. We never do this for what happens out of view, or for news that's only just broken.
• Opponents who agree do count. When two parties on opposing sides of a matter say the same thing, neither has a reason to back the other. The judge has to state why they're opposed, from the evidence on record, and we show that reason.
• Explanations have to trace back. Each statement we publish cites its evidence. If it names someone or gives a number that isn't in the claim or its evidence, we remove it, because it came from the AI's memory and not from the research.
• Not proved isn't shown false. Both are Refuted, but each claim says which, so you can tell a claim the record contradicts from one nobody backed up.
• Old news can't confirm new events. Coverage published before an event doesn't count as evidence it happened.
• If the evidence against weighs as much as the evidence for, the claim hasn't been proved, so it's Refuted.
• One false part makes the whole thing false. A small lie inside a true story is how false stories spread, so if any part is refuted, the post is Refuted, however true the rest is. We name the part that failed.
• Confirmed means every part. A post is Confirmed only when every part of it is. We test every claim in what you submit, however many there are, and list anything we couldn't. We can't yet show that we covered every detail of a linked article, and we say so there.

Rating an account

We go through the account's recent posts in order, newest first, take one claim from each (its main fact, without incidental details), and test each one the same way. We don't pick and choose. Only claims the account itself asserts count, so passing on or debunking a rumour isn't held against it. Posts that readers submit never count, so nobody can sink an account by submitting its worst moments.

• Trusted

  A tenth or less of its recent checkable claims were refuted.

• Mixed record

  More than a tenth of its recent checkable claims were refuted, but less than a quarter. Check each post on its own.

• Low trust

  A quarter or more of its recent checkable claims were refuted.

In exact terms: Low trust means 25% or more of its claims were refuted, Trusted means 10% or fewer, and Mixed record is in between. Every claim it makes counts, because each one is Confirmed or Refuted. An account whose recent posts make no factual claim of its own gets no rating. Accounts reviewed before November 2026 are reviewed again the next time they're opened.

A trust rating is a summary of recent claims, not a judgement on anyone's honesty or intentions, and a paid check-mark on X has no effect on it.

Showing who an account is tied to

We show affiliations with governments, militaries, intelligence services, political parties, state media, lobbies, armed groups, companies, NGOs and others. A wrong label could put someone in danger, so we report only what's documented:

• Labelled by X: X's own government verification or affiliate badge.
• Stated by the account: the exact words in its own bio.
• Public record: official registers and Wikidata, a public, community-maintained database.
• Reported by news media: reputable reporting that states it directly, with the exact words checked on the page.

We never work out an affiliation from what someone posts, who they follow, their language or their location. Having one doesn't change a trust rating: a government account earns its rating like everyone else.

Who we don't rate

We rate public-interest accounts only: governments, organisations, public figures, and accounts with over 10,000 followers. Private people get no rating and no affiliation label, though you can still fact-check what they post. Private accounts are never processed, and accounts that describe themselves as parody or satire aren't treated as making factual claims.

For individuals, we show only public roles. We never show anyone's religion, ethnicity, health or sexuality, or the private politics of anyone not in public political life.

Analysing a narrative

A narrative starts from a topic and a country. We read what was posted about it on X over the window chosen, in the languages people there use, and find the storylines: what each says happened, who it blames, and whose interests it would serve. Every claim a storyline rests on is checked like any other claim, and the footage it uses is analysed the same way as a post's video.

Each storyline gets one conclusion, its worst problem: Fabricated, False footage, Exaggerated, Amplified, or No manipulation found. It's worked out several times over, independently, and we show how many times each conclusion came up. We never say who is behind a storyline, or that it was deliberate, unless a checked source says so. Where authors are comes only from what they typed in their profiles, and where an outlet is based isn't evidence of what it intended.

Same answer on every plan

Everyone sees the same verdict. Paying gets you more detail, volume and tools, never a different answer. The one exception is a Pro deep check, which researches further and uses that person's own source lists. It's private and clearly labelled.

A result belongs to the reader who paid for it. Anyone else sees its verdict and summary, and can see the rest by spending one of their own checks, or one of their narratives for a narrative.

When we get it wrong

Every result has a link to ask for a review, and you don't need an account to use it. A person reads each request. If it's upheld, the label comes down straight away for everyone.