How TruthLens Checks Facts

Effective date: 8 September 2026 · Methodology version 1.7

This page explains, in plain language, how TruthLens turns a piece of text, a video, or an image into a verdict backed by sources. It is written for readers, journalists, and researchers who want to know what is actually happening behind a result — not marketing language. Where a claim below can be independently checked in our Terms of Use or Privacy Policy, we link to the exact clause rather than restate it differently.

An analysis tool, not a final verdict. Every result TruthLens produces — verdicts, confidence scores, explanations, source lists — is generated automatically by AI models reasoning over public evidence. It is a research and information tool, not an established fact and not a final verdict (see Terms of Use, clause 6.1). TruthLens is not affiliated with, and is not a verified signatory of, the International Fact-Checking Network (IFCN); the principles on this page are informed by IFCN's published Code of Principles, applied to an automated system.

1. How a check works

A check runs in stages. Every stage is designed around one rule: no verdict without supporting evidence.

1.1. Extraction. The submitted text (or the transcript of a video, or the text read from an image) is broken into individual, checkable factual statements — named events, statistics, and claims presented as fact. Opinions, predictions, and rhetorical questions are not extracted. The overall piece is also classified as news, opinion, or satire.

1.2. Deduplication and prioritization. Repeated or near-identical claims are merged. For long material, the claims judged most consequential to a reader are checked first; a fixed number of the most significant claims per submission are checked in full — this keeps checks fast and keeps the checking budget on what matters, not on incidental detail.

1.3. Matching against published fact-checks. Before running our own search, we check whether a fact-checking organization that is a verified signatory of the IFCN Code of Principles has already published a review of the same claim. A match is accepted only when the matched claim is confirmed — by exact wording or by a dedicated verification step — to be about the same assertion, not merely the same topic. When it matches, that organization's own conclusion is used instead of ours.

1.4. Evidence search. For claims without an existing fact-check, TruthLens searches the open web for supporting or contradicting evidence, excluding social-media posts as primary sources. Claims about scientific or health topics are additionally checked against scholarly and biomedical literature databases (OpenAlex, Semantic Scholar, and PubMed — these index published research broadly, not only work that has completed peer review); claims about a country-level statistic (GDP, inflation, population, unemployment, life expectancy, and similar indicators) are checked against official figures from the World Bank. If a page has been removed or is temporarily unreachable, we attempt to read the last publicly archived copy via the Internet Archive's Wayback Machine before giving up.

1.5. Grounded verdict. An AI model judges the claim strictly against the evidence collected in the previous step — never against its own general knowledge. If the evidence is insufficient or contradictory, the verdict is Not confirmed, not a guess. The model must also state what the claim asserts and what the evidence asserts. Evidence about a neighbouring statement — a bill introduced rather than a law passed, a person charged rather than convicted — counts as neither confirmation nor contradiction. A verdict that contradicts the model's own comparison is discarded in code: both a confirmation and a refutation become Not confirmed, with a note on what the sources actually address. Evidence about a different event or assertion is neither confirmation nor refutation. A source that gives only a bound — “over $55 billion” — does not establish an exact figure inside it: such a claim is Not confirmed, not Misleading and not False, and the note says what the sources do give. A report about an earlier state of affairs may be insufficient to refute a claim about the present: we examine available publication dates, dates of the reported data, and explicit updates in the retrieved excerpts; an undated source is assumed neither old nor current. When the available material does not establish a sufficiently current basis for refutation, the negative verdict is withdrawn to Not confirmed and the note explains the temporal limitation. Misleading requires evidence of a real fact and of how the claim distorts it; missing or unrelated evidence alone is insufficient.

1.6. Source-trust weighting. Each evidence source is weighted by trust in a graph of known media and publishers. The model performing the grounded verdict (1.5) is instructed to never call a claim False or Misleading on the strength of low-trust sources alone, and to defer to a high-trust source when a high-trust and a low-trust source disagree. Separately — and enforced in code, not left to the model's judgment — a False or Misleading verdict is not shown with high confidence unless it is backed by at least two independent sources; reprints of the same wire story on different domains count as one. Agencies and offices of the same government count as one source in the same way.

1.7. Independent second opinion. Most False or Misleading verdicts are separately re-examined by a second AI model given the same evidence, up to a fixed budget per check. If the two disagree, the verdict is shown with reduced confidence. A False verdict that a second model of comparable strength does not support is withdrawn to Not confirmed rather than presented as a refutation — unless the second model itself establishes how the claim distorts a real fact, in which case the claim is shown as Misleading. A second opinion that finds a claim unresolved does not agree with a Misleading verdict. When disagreement prevents a supported conclusion, we report the claim as unresolved rather than manufacture a weaker accusation. A verdict that already matches a published fact-check (1.3) skips this step — a human fact-checking organization's own review already stands behind it.

1.8. Aggregation. When a submission contains several claims, the overall result reflects all of them: a single false statement is not enough to brand an otherwise accurate piece as fake, and a single true statement does not redeem a piece built mainly on false ones. The overall score summarizes the claims selected for checking; it is not a probability that the entire submission is true. Unresolved claims do not count as confirmation and cannot dilute the share of claims found false. Lower confidence in a positive finding reduces its contribution; lower confidence alone never increases the overall score. A safeguard that withdraws an unreliable verdict reduces coverage but does not, by itself, impose an additional negative judgment.

2. Verdict scale

Verdicts are the same across the app, the browser extension, and this website. Two scales apply: one for the overall submission, one for each individual claim inside it.

Overall result

Individual claims

Two claims are the author's own statement about themselves (for example, "I started this channel a month ago"). These are not checked against outside sources at all and are always marked Not confirmed, on principle — no external source could confirm or deny them.

Misleading, more precisely

Where the underlying reason fits one of two well-known patterns, we say so rather than leaving a bare "Misleading" label:

3. How we judge a source

Not all evidence is weighted equally. Every source is considered on four criteria: authority (the organization's standing and role on the topic), transparency (open methodology, clear data), recency (freshness relative to the claim being checked), and independence (absence of conflicts of interest). In practice this becomes a three-tier trust level — high, normal, or low — built from a graph of known media outlets and a short list of known low-quality domains (content farms, unmoderated blogs). The model is instructed to treat a contradiction found only on a low-trust source as insufficient, by itself, grounds for a False or Misleading verdict, and to follow the high-trust side when a high-trust and a low-trust source disagree — see 1.6 for the one part of this weighting that is enforced in code rather than left to the model.

4. What TruthLens does not do

• It does not produce a legal finding, a medical or financial recommendation, or an official determination of any kind (Terms of Use, clause 6.4).
• It does not check every claim in a long submission — by design, it prioritizes the most significant ones (see 1.2).
• It depends on evidence that is public, indexed, and available online at the time of the check; claims about purely private matters, or about events with no public trace yet, will typically come back Not confirmed rather than resolved.
• It relies in part on free, third-party infrastructure (open web search, public archives, open scientific and statistical databases). When one of these is temporarily unavailable, TruthLens degrades to the next available source rather than failing outright — but a temporary outage can occasionally mean less evidence than usual for a given check.
• Explanations and other free-text output are written directly in the language of the app you are using — 35 of the 36 languages the interface supports — rather than machine-translated after the fact. Danish is the single exception: our own measurement found the model drifting into Norwegian on claims about Norway, so a Danish reader gets the interface in Danish and the explanations in English until that is fixed. If you read the app in a language outside the 36, both the interface and the output fall back to English. This follows the language you read in, not the language of the material: a German article checked in a Thai app comes back in Thai. The standard messages the service writes itself — for example the note shown when a published fact-check already covers your claim, or the explanation of why a submission produced nothing to check — now cover the same 36 languages, but they are machine-translated drafts that have not yet been read by a native speaker of every language.

5. Correcting a result

If you believe a specific result is wrong, or a result concerns you or your organization and you consider it incorrect or harmful, email support@truthlens.wiki with the exact claim checked and the substance of your objection. We review reasoned notices within a reasonable time — as a rule, within 14 days — and, where justified, correct or remove the disputed result and take it into account going forward (Terms of Use, clause 6.6).

6. Changes to this methodology

Material changes to how checks are performed will be dated here, in plain language, at the same time they ship.

• 8 Sep 2026Version 1.7 — 1.5: a source that gives only a bound (“over $55 billion”) no longer establishes an exact figure inside it, and a negative verdict resting only on sources about an earlier state of affairs is withdrawn to Not confirmed with a note on the temporal limitation; a refutation that contradicts the model's own comparison of predicates is now withdrawn to Not confirmed instead of being shown as Misleading; 1.7: a False verdict that a comparably strong second model does not support is withdrawn to Not confirmed unless the second model itself establishes a distortion; 1.8 and 2: the overall score no longer lets unresolved claims dilute the share of claims found false, lower confidence never raises the score, and the overall labels are defined by what was resolved, with coverage reported separately.
• 5 Sep 2026Version 1.6 — 1.5: the verdict model must now compare what the claim asserts with what the evidence asserts, and a verdict that contradicts its own comparison is discarded in code; 1.6: agencies of the same government count as one source; 1.7: a refutation that a comparably strong second model does not support is downgraded to Misleading instead of being shown with lower confidence only.
• 15 Aug 2026Version 1.5 — named the one language where free-text output is not written in the language of the app. Our own measurement of output language across all 36 found Danish requests coming back in Norwegian Bokmål on claims about Norway, reproduced over three independent runs, so Danish explanations are written in English while the interface stays Danish. The previous wording implied output in all 36 without exception.
• 15 Aug 2026Version 1.4 — corrected the output-language statement in 4 once more: it described the language of the submitted material, when the output actually follows the language of the app you read in. A German article checked in a Thai app is written back in Thai; English is the fallback only when the app itself runs in a language we do not support.
• 15 Aug 2026Version 1.3 — the standard messages written by the service itself (the fact-check match note and the explanations of an empty result) now exist in all 36 languages rather than five; the statement in 4 that some of them appear in English has been replaced with the fact that they are machine-translated drafts pending native review.
• 15 Aug 2026Version 1.2 — corrected the output-language statement in 4: free-text output is written in the language of the app across 36 languages, not five, with English used for anything outside that set and for a few standard messages that are not yet translated.
• 28 Jul 2026Version 1.1 — precision pass on 1.4 (literature databases are not exclusively peer-reviewed; the statistics list is not open-ended), 1.6 (separated the AI-instructed source-trust weighting from the code-enforced independent-sources rule), 1.7 (the second opinion applies to most, not every, False/Misleading verdict, and is skipped for matched fact-checks), and the Fake row in the verdict table (replaced "predominantly" with a description matching the actual threshold).
• 28 Jul 2026Initial publication (version 1.0).