Veritap's algorithm
How Veritap decides whether a statement is true, false… or manipulative.
No black box. Here, in plain terms, is how Veritap pits several AIs against each other, compares sources and computes its scores, both when listening live and on a piece of content you hand it.
Several AIs, never just one
Veritap automates what everyone naturally does to check a claim: search the Internet for sources, then ask AIs for their opinion. Veritap first searches the web for sources, then sends each statement, along with those sources, in parallel to a panel of independent frontier AIs: Grok, Gemini, Mistral and Claude. Each one returns its verdict, citing the sources that back it. Veritap then confronts these opinions: unanimity strengthens the conclusion, disagreement is shown rather than hidden.
Web research first
Veritap queries the web at the moment of analysis and hands the AIs fresh sources, not a frozen memory. Recent facts are taken into account.
Evidence wins
A verdict backed by sources counts up to 3 times more than a mere opinion with no reference. One AI that actually checked beats two that simply believe.
Disagreement is visible
When the AIs diverge, Veritap says so ("2/3 agree") instead of deciding in silence. You see the level of consensus.
The four verdicts
Every statement gets a colour-coded badge, along with a confidence level and the number of AIs in agreement. And since Phase 10, Veritap detects irony and second-degree meaning: an ironic statement is not judged at face value — the verdict bears on what the author actually argues.
Confirmed by reliable, concordant sources.
Partly accurate, but misleading, incomplete or taken out of context.
Contradicted by reliable sources.
No public source settles it. Important: an unverifiable statement never counts against the speaker. "Impossible to check" is not the same as "a lie".
The lie score, from 0 to 100
The needle only rises on what is factually false. Only "false" verdicts (full weight) and "partially true" ones (half weight) push it up, weighted by the confidence and importance of the statement. A single big lie is enough to move the BullshitMeter. True or unverifiable statements carry no weight at all.
Presumption of good faith: as long as nothing is proven false, the score stays low.
A second radar, independent of the facts
You can string together ten true statements and still build a misleading argument. That's why Veritap measures a second axis, in purple: rhetorical manipulation. Fallacies (straw man, ad hominem, false dilemma), misleading devices (whataboutism, loaded language) and emotional pressure (fear, urgency, guilt-tripping) are spotted and explained.
Its own scale
The manipulation score is computed separately: severity of the device × certainty of the detection. Several devices push the purple needle up.
Objective detection
Veritap identifies manipulation mechanisms from recognised rhetorical patterns. The analysis is systematic and free of human bias.
Educational
Every detected device is named and explained: Veritap doesn't just flag, it sharpens the mind.
Live or on content: the same engine
Veritap listens to the speech picked up by your smartphone, glasses or browser, and shows verdicts and scores as the conversation unfolds.
Paste a video link: Veritap transcribes and analyses it end to end, with exactly the same verdicts, sources and scores.
Why you can trust the verdict
Sources, always
Every factual verdict cites public references, clickable and verifiable by you.
The power of the face-off
No single AI decides. Cross-checking opinions neutralises the error or bias of any isolated model.
Two measures, not one
Truthfulness and manipulation are scored separately: you know whether a statement is false, manipulative, both, or neither.
Honest about its limits
When a fact can't be established, Veritap admits it ("unverifiable") instead of inventing certainty.
The sources Veritap consults
Depending on what a claim is about, Veritap grounds its AIs in open sources — web search and public databases. Our thanks to the projects that make verification possible:
Web search (SearXNG)
self-hosted open-source metasearch engine — cited pages remain their publishers' own
Wikipedia & Wikidata
free encyclopedia and knowledge base (public source profiles)
Google Fact Check Tools
fact-checks already published by newsrooms (AFP Factuel, Les Décodeurs, Snopes…)
World Bank
official statistics (inflation, unemployment, GDP, population…)
Europe PMC
biomedical literature
OpenAlex
open-access scholarly publications
Crossref
scholarly publication metadata, including retractions
GDELT
global news monitoring
Eurostat
official European statistics (harmonised data from INSEE and national institutes)
These sources provide context: none of them decides a verdict on its own, and every result card cites the references actually used.