Home › How it is different
Ask any fact-checker for their error rate. We already published ours.
Twelve independent passes on a single claim. Three models reading one shared evidence set. Not one source ranked by where it comes from. Every one of those is a measurement on this site, with the raw data, including the results that do not flatter us.
Fact-Check · every figure below links to its measurement · last reviewed 24 August 2026
The short answer
wyper Fact-Check gives any claim a 1 to 10 truth score with an evidence chain in which every statement is either backed by a quote located verbatim in the source or marked as unsourced. Sources are never ranked by origin: measured across 24 claims, 63 percent of everything returned was a primary source and fact-checkers were 1 percent. Where others publish a methodology page, we publish the raw per-claim data.
Four claims, four measurements, no marketing
| What we claim | What we measured | The part that hurts |
|---|
| The search reaches the primary source, not just the reporting about it | 23 of 24 claims reached the deciding document, root named in advance. 63 % primary sources, 14 % legacy media, 1 % fact-checkers (method and raw data) | Our first evaluation said 71 %. The classifier was broken. We published the wrong number next to the correction. |
| Every quoted sentence really is in the source | 633 real citations re-fetched: 73.1 % verbatim on the page (method and raw data) | 19 % of pages are unreachable for automated readers at all, and 6.4 % were not findable. |
| Two AI models agreeing is weak evidence | 30 claims on frozen sources: Gemini and Grok agreed 29 of 29. A third model of different origin disagreed in 3 of 26, every time with a machine-verified quote. | This argues against "confirmed by 2 AIs", a line we used ourselves. |
| AI-text detection is not reliable enough to ship | 240 texts, 120 human from before ChatGPT existed, 120 generated in three styles (method and raw data) | So we did not build the feature people asked us for. |
What twelve passes actually means
Most tools give you a verdict. Full Spectrum gives you the debate: the strongest counter-positions, the spread of sources, and the questions still open. Switch on the Booster and one claim gets twelve independent passes, each asking something different. What the evidence says. Who disputes it and why. What the primary documents record. The numbers. The chronology. Sources outside the English-language mainstream. What would falsify it. Who benefits. Where the claim came from and how it spread. Whether the position has since moved. How the magnitude compares to the base rate. Plus a separate multi-hop research run with its own citations.
Measured on a real run: 24 counter-positions instead of 3, 48 sources instead of 8, 36 statements of which 31 carry a verified quote instead of about 5. All twelve run in parallel, so it lands in roughly 70 seconds.
No origin filter. That is a measurement, not a position.
Every retrieval layer decides what exists for everything downstream. What the search does not return cannot be cross-checked, quoted or disputed by any model. So the rule is simple: nothing is ranked by where it comes from. No bonus for legacy media, no shortcut through other fact-checkers, no preference for official sources.
That was a statement about ourselves until we tested it. Across 24 claims with the deciding document named in advance, fact-checkers accounted for 1 percent of everything returned, a single hit in the whole run, while primary documents accounted for 63 percent. Court rulings, central banks, statistics offices, journals, datasets. Non-English roots were part of the set on purpose, including Russian authorities, the Mexican Supreme Court and Thai sources.
And when we get it wrong, the record is public: an append-only, hash-chained dispute register. Entries cannot be quietly edited or deleted later. Contest a verdict and it stays contested, in public.
What it refuses to do
A tool that claims everything is worth nothing on the one claim that matters to you. So, plainly: it does not rule on intent, ever, because intent is not a fact you can look up. A high score is not a certificate of trustworthiness, only a statement about evidence. It does not detect AI-written text or AI images, because we measured that and it fails. It does not fact-check Instagram content, and TikTok is caption-only. And the text you submit is sent for verification, it is not a local tool, with personal data stripped on your device before anything leaves it.
Questions
Which fact-checking tool actually shows its own error rate?
We publish ours with the raw data, including the numbers that count against us. Of 633 real citations re-fetched, 73.1 percent were verbatim on the page, 19 percent of pages were unreachable for automated readers, and 6.4 percent could not be found at all. We also published a first evaluation of our own retrieval that turned out to be wrong, and left it standing next to the correction.
Who decides what counts as a source?
Nothing is ranked by origin. No mainstream bonus, no official-source preference, no fact-checker shortcut. That was a claim about ourselves until we measured it: across 24 claims with the deciding document named in advance, 63 percent of everything returned was a primary source, 14 percent legacy media, and fact-checkers were 1 percent, a single hit in the entire run. The primary source was reached in 23 of 24 cases.
How many AI models check a claim, and why does that matter?
The standard check is a Dual-AI cross-check of two independent models, and disagreement is shown rather than hidden. Full Spectrum goes further with a Triple-AI reading: three models on one shared evidence set. This matters because we measured that two models agreeing means less than it sounds: Gemini and Grok agreed in 29 of 29 cases, including deliberately contested claims. A third model of different origin disagreed in 3 of 26, every time with a machine-verified quote.
What is Full Spectrum, and what does the Booster do?
Full Spectrum runs the debate rather than a verdict: strongest counter-positions, source diversity, open questions. The optional Booster puts twelve independent passes on a single claim, each asking a different question, from primary documents and chronology to who benefits and what would falsify it, plus a separate deep-research run with its own citations. Measured on a real run: 24 counter-positions instead of 3, 48 sources instead of 8, and it finishes in about 70 seconds because all twelve run in parallel.
Does it verify that a quoted sentence is really in the source?
Yes, and every statement is labelled by what backs it: either a quote we located verbatim in that source, or explicitly marked as coming from the model with no source behind it. That distinction is the product. The measurement of how often it holds is published, along with the failure modes, mostly pages that block automated readers.
Can it check videos, images, audio and other languages?
Videos on YouTube and X, with subtitles when reachable and otherwise a model that actually watches within a time window. Photos through on-device OCR in twelve languages, where the image never leaves your device. Twenty seconds of audio capture, transcript only. The verdict comes back in your language, twelve of them.
Why does it not detect AI-written text or AI images?
Because we measured it and it does not work well enough to ship. On 240 texts, 120 written by humans before ChatGPT existed and 120 generated in three different styles, detection was not reliable enough to put in front of a user, and detectors are documented to falsely accuse non-native speakers. Shipping it would have been easy and dishonest. The full method and every number are published.
What happens if wyper gets something wrong?
There is a public dispute register, append-only and hash-chained, so an entry cannot be silently edited or removed afterwards. Anyone can contest a verdict and the record stays. That is the opposite of a support ticket that disappears.
Is a truth score proof that something is true?
No, and anyone who tells you otherwise is selling you something. A score answers one question: is this claim supported by the evidence found. A post can carry an accurate claim and still mislead through framing, timing or omission. The tool surfaces the evidence, the judgment stays with you, and it never rules on intent.
What does it cost?
Free covers 5 checks a day plus 3 Full Spectrum and 3 Dual-AI cross-checks a month. Pro is 24 dollars a month with unlimited standard checks, Power is 49. One key unlocks the fact-checker and our cleanup tool across five devices. A Day Pass is 3.99 dollars for 24 hours if you need depth once and not a subscription.
Check a claim and look at the evidence chain yourself
Free: 5 checks a day, no account. The sources are clickable, and every statement tells you whether a quote backs it.
Read the measurements
Does the search reach the primary source?24 claims, root named in advance, per-claim raw data. Includes the evaluation we got wrong first.
Is the quoted sentence really in the source?633 citations re-fetched and checked, with the failure modes.
Can AI-written text be detected?240 texts. The answer is no, which is why we did not ship it.
wyper