A FAKTS check is an AI-assisted assessment. This page explains, in plain language, exactly how a claim gets a verdict, what the numbers mean, and where the method has limits. Nothing here is hidden behind marketing.
1. Claims are extracted
You give FAKTS a piece of content: a pasted claim, a document, a URL, a media transcript, or a whole site. FAKTS pulls out the individual factual statements. A sentence like "The Eiffel Tower is 324 metres tall and receives 7 million visitors a year" becomes two claims, checked independently. Rhetorical questions, opinions, and pure predictions are ignored; only claims that can be checked against sources are put forward.
2. Each claim goes to the consilium
Every claim is sent to a panel of independent models, the consilium. Panelists are anonymous in the results so no single provider's brand shapes how you read a verdict. Some panelists are web-grounded, meaning they can pull live sources at the moment they answer; others reason from what they were trained on. Both kinds respond in a common structured format so their verdicts can be compared.
3. Web-grounded evidence is weighted more
When a panelist backs its verdict with a fresh, retrievable source, that answer counts more in the consensus than one that reasons from memory alone. This does not mean the ungrounded panelist is discarded; it means the tie goes to whichever verdict has verifiable evidence behind it. Calibration weights are tuned against a benchmark set of claims with known answers, so the weighting is empirical, not editorial. Claims are additionally checked against an independent search index; sources it returns are listed alongside panelist citations.
4. Disagreement is shown, not hidden
Panelists routinely disagree. When they do, FAKTS surfaces the dissent inside the result card: which panelist argued for a different verdict, and, when useful, the argument they made. A verdict without visible dissent is a verdict with none; a verdict with visible dissent is one you can still trust, but that you can now inspect and challenge.
5. Calibrated consensus produces the verdict
The weighted votes are combined into one verdict per claim. A confidence number expresses how strong that consensus was, on a 0 to 100 scale. High confidence means the panel converged and its sources agreed. Low confidence means the panel was split, or the evidence was thin, and the verdict should be read with more care.
The five verdicts, defined
- Verified. The claim is supported by strong, live, independently retrievable sources. Reasonable readers should treat it as accurate.
- Emerging. The claim is supported by preliminary evidence but has not yet reached broad recognition. It may be correct today and settled tomorrow; read it with context.
- Misleading. The claim contains a real fact but frames it in a way that leads a normal reader to the wrong conclusion. A correction is provided.
- False. The claim is contradicted by strong sources. A corrected version is provided.
- Unverifiable. The claim cannot be resolved with the sources available at check time. This is a real outcome, not a failure. It is a signal that a human, or a specialist, should look further before accepting it.
Site scores and seal bands
A site score is the aggregate of the verdicts on that site's claims. Verified counts fully, emerging counts partly, unverifiable counts as a small credit for honesty, misleading counts little, and false counts as zero. The score is on a 0 to 100 scale.
- Verified seal. Score of 80 or above, zero false claims, at most two misleading claims.
- Provisional seal. Score between 60 and 79, and zero false claims.
- No seal. Anything below that; the report is still public, but the site does not carry a seal.
Seals are living: they are recomputed whenever the underlying report is re-run, and they can be revoked when new evidence overturns the earlier assessment. Revocation is public.
What the confidence number means
Confidence is not certainty. It expresses how consistently the consilium reached the verdict under the current evidence. A verified claim with confidence 92 is stronger than a verified claim with confidence 61, but both are verdicts you can inspect. Confidence updates when a claim is re-checked and new sources are available.
Human review, when it matters
Some claims deserve a human editor: medical, legal, regulated marketing, and anything with a legal consequence. The Comply layer produces a signed compliance record; the Experts add-on adds a named subject-matter reviewer whose sign-off is attached to the result. Both are visible on the certificate and on the shareable card.
Limitations, honestly
FAKTS is an AI-assisted assessment against the sources available at the moment of the check. It is not an oracle and it is not a guarantee. Concretely:
- Sources can change. A verified claim may become misleading if the source is later corrected or the underlying reality shifts. Re-verifying is one click.
- Unverifiable is a real outcome. It means the evidence available right now is not strong enough to rule. It does not mean the claim is false, and it does not mean the check failed.
- Very fresh events sit awkwardly with any method that leans on sources: emerging is the correct verdict for a claim whose evidence is still forming.
- Providers change models over time. Calibration weights are re-fit against the benchmark so the aggregate stays honest, but any single verdict is a snapshot in time.
- No fact-checking system, human or machine, is free of error. This is why every verdict links its sources, dissent is shown, and a corrections route is offered on every surface.
You can always ask for a re-check, request an expert, or file a correction. That is part of the method, not an afterthought.