How Verifyd works
Verifyd is an evidence interface, not a truth detector. It never decides on its own whether something is true. Each word it shows comes from evidence published by someone accountable, and every word links to that evidence so you can check it yourself.
Three separate questions
Verifyd keeps three kinds of evidence apart and never merges them into one truth score:
- Claims: what published fact-checks say about a factual statement.
- Origin (provenance): what is known about where a photo came from and whether it was changed.
- Context: labels that are not verdicts, such as OPINION, OLD STORY, CORRECTED, STATE MEDIA and PUBLIC MEDIA.
A real photo can carry a false caption, and a true statement can be UNCHECKED because nobody has fact-checked it. One word is shown first; the separate answers are listed underneath.
The words, and the evidence each one needs
| Word | Only when |
|---|---|
| TRUTH | A published fact-check rated the same claim true. |
| FALSE | A published fact-check rated the same claim false. |
| CONFIRMED | No fact-checker has rated the claim, and at least two independent, established news or official sources directly confirm it. Each source is linked with the passage it says, quoted word for word. |
| CONTRADICTED | No fact-checker has rated the claim, and at least two independent, established news or official sources directly report it is untrue. Each source is linked and quoted. |
| DISPUTED | A fact-check rated it partly true, misleading, missing context, unproven or unsupported, or fact-checkers genuinely disagree. |
| REAL | The photo carries valid Content Credentials from an issuer on the trust list, and it has not changed since it was signed. This proves its origin, not that the scene shows the whole story. |
| AI | The photo's own Content Credentials or metadata declare that it was made by AI. |
| LIKELY AI | An AI detector rates it very likely AI-made. A detector score is a probability, so it is a warning, never the AI verdict on its own. |
| ALTERED | The photo's credentials no longer match its pixels, or it declares an AI edit. |
| OLD | The same picture appeared online earlier than the page presents it. Old pictures are not deceptive in themselves. |
| FAKE SITE | The address is built to deceive (look-alike characters, one letter away from a famous site, or the real address placed inside another), or the page presents itself as the brand it imitates. |
| LOOK-ALIKE | The address uses a famous name but nothing else confirms impersonation. A warning, not a verdict. |
| UNCHECKED | No eligible evidence either way. Not a verdict. Most of the internet is unchecked. |
Where claim evidence comes from
Fact-checks come from published ClaimReview records (the open schema.org standard fact-checkers use to label their work). Verifyd keeps its own index of them, built from three independent sources: the open ClaimReview data feed (around 100,000 fact-checks, read daily), the ClaimReview markup on fact-checkers' own websites (read hourly), and the Google Fact Check Tools search, used alongside when available. If any one source is unavailable, the others still answer. Verifyd's own search finds candidates by the claim's words, weighting rare words more, matching word forms and common synonyms (COVID-19 and coronavirus), and never letting a different number match. The fact-checks are written by fact-checking organisations such as AFP Fact Check, PolitiFact, FactCheck.org, Snopes, Full Fact, USA Today, Reuters and many others in several languages. Verifyd shows each fact-check with its publisher, its exact rating and its date, and links to it. Verifyd does not write fact-checks and never presents a fact-check as its own.
When no fact-checker has covered a claim
Verifyd then searches the live web and reads what established news organisations, official and public bodies, encyclopedias and scientific sources report. It counts only passages those pages actually contain, quoted word for word, from at least two different sites. Social media, forums, video sites and blogs never count. Opinions, predictions and claims the sources do not settle stay UNCHECKED. This happens automatically after the quick fact-check search; the result says which kind of evidence it rests on.
Matching a fact-check to a claim
- Find candidates. A keyword search returns fact-checks that may be about the claim. English is searched first; other languages are searched when needed.
- Same words, same answer. When every candidate rated almost exactly the same words, with every number identical, it counts at once.
- Otherwise, strict matching, then a second look. An AI model, run with fixed settings so the same input gets the same answer, decides which candidates rate the same assertion. It follows written rules: the same subject, action, number, place and direction; for facts that change over time, the same period. A fact-check of what a document, video or person said about a claim does not rate the claim itself. The model never judges the claim. Each fact-check it counts is then checked again on its own, one pair at a time: the same statement, the opposite, or a different one. Only the same statement can decide the word.
- Related, not counted. Candidates that rate a different claim are still shown, labelled as related, and never decide the word.
Reading a fact-checker's rating
Each rating is kept exactly as published and also read into one of four groups. The opening words of a rating decide it ("Partly false. The video…" is partly false, not false).
- False: false, fake, fabricated, hoax, incorrect, inaccurate, not true, no evidence, unfounded, baseless, doctored, Four Pinocchios, and the same in Spanish, Portuguese, French, German and Italian (falso, faux, falsch…).
- Mixed (DISPUTED): partly or partially false or true, half true, mostly false, misleading, out of context, missing context, exaggerated, unproven, unsupported, disputed, two or three Pinocchios (engañoso, trompeur, irreführend…).
- True: true, correct, accurate, mostly true, verified (verdadero, vrai, richtig…).
- Anything else (for example "Satire" or "Explainer") gives no word.
When fact-checkers disagree
If fact-checks of the same claim say both false and true, the claim is DISPUTED and both are shown. When false ratings outnumber true ones at least two to one, the claim stays FALSE, and the true rating is still listed.
Photos and Content Credentials
Verifyd reads Content Credentials (the C2PA standard) with the open-source c2pa library rather than its own cryptography. It checks the signature, whether the credentials still match the pixels, the issuer against the official C2PA conformance trust list (which includes camera and phone makers such as Google and Samsung, whose phones can sign every photo) and the Content Authenticity trust list, edit history and AI declarations, including the labels generators add under the EU AI Act. Missing credentials are normal: most photos have none, and social platforms remove them. So no credentials never means fake. Verifyd also keeps a fingerprint of each picture to notice when an old picture is reused.
Speeches
Speeches come from official transcripts or captions. Each factual claim is shown with the speaker's exact words when those words are found in the transcript; otherwise it is marked "In summary" and never shown in quotation marks. A speech is never given one overall verdict: each claim carries its own word, and the totals are shown. Every speech is checked again daily as fact-checkers publish. The same rules apply to every speaker, party and country.
News outlets
STATE MEDIA marks outlets whose government controls what they publish. PUBLIC MEDIA marks outlets that are publicly funded and editorially independent by charter. Both are taken from what outlets state and from public records, applied to every country the same way. They are context, not a judgement of any story.
Corrections
When a fact-checker publishes or revises a rating, the next check uses it; speeches are rechecked daily and saved answers expire within hours. If you think a word is wrong, write to hello@planckchron.com with the link. We will check the evidence and correct the rule or the result, and say what changed.
Known limitations
- Most claims have no published fact-check, especially new ones, so most results are UNCHECKED.
- The AI detector and reverse-image search are not switched on at present, so LIKELY AI and OLD are rare.
- Comparing a claim against news coverage needs a web search service that is not switched on at present.
- Matching uses an AI model under strict rules; it can still make mistakes, which is why every word links to its evidence.
- Fact-check coverage is strongest in English and uneven across countries and languages.
- Verifyd has not been independently audited. It publishes its own test results, case by case, on the test results page; they are self-tested and say so.
Privacy
Nothing is sent before you agree. Verifyd shows no advertising, sells no data and builds no profiles. See the privacy policy.