Verifying a Claim Is Easy. Opening a New Tab Is the Hard Part
You are reading an article, or watching an interview, and a sentence stops you. A number that seems too round. A quote attributed to someone who probably never said it. You know exactly how to check it: open a tab, search, read two or three sources, come back.
Almost nobody does. Not because the check is hard, but because the check happens somewhere else. By the time you would have finished, you have lost the paragraph you were in, and the doubt has quietly turned into acceptance.
DeepVerify is a Chrome extension that removes that trip. It runs on the page you are already on, checks claims where they appear, and puts the result next to the sentence that raised the question.
A Verdict Is Not an Answer
Most automated fact-checking ends at a label. True. False. Misleading. That label is not a conclusion, it is a new thing to trust. If a system tells you a claim is false and shows you nothing else, it has not resolved your doubt. It has only moved it from the article to the checker.
So every DeepVerify verdict arrives with the work attached:
- ●Verdict: true, false, or uncertain
- ●Confidence score: how strongly the evidence supports that call
- ●Sources checked: the actual documents used, not a list of topically related links
- ●Reasoning: why those sources led to that verdict
- ●Recheck: run the same claim again later, when the evidence may have changed
- ●Go deeper: post the claim on factagora.ai when a verdict is not enough
The point of the card is that you can disagree with it. You can open the sources and decide the checker read them wrong. A verdict you cannot audit is a verdict you have to take on faith, which is the problem we started with.
Three Ways a Claim Reaches You
Claims do not arrive in one format, so DeepVerify does not have one mode.
- ●Highlight it. Select any sentence on any page and get a verdict on the spot. This is the narrowest case, and the most common one: you have one specific doubt and you want it answered without losing your place.
- ●Let the page scan itself. Auto-detect reads the page, pulls out the factual claims, and marks the ones worth a closer look right where they appear. It answers a different question than highlighting does. Not "is this sentence true" but "what in here should I be questioning at all."
- ●Listen as it happens. Live audio captures your microphone or the audio playing in the current tab, transcribes it in real time, and checks claims the moment they are spoken. Livestreams, webinars, and calls have no pause button for looking things up, so the verdict has to arrive while the conversation is still going.
On YouTube, these combine into one view. A card sits beside the video with every checked claim in timeline order. Click a claim and the video jumps to the second it was said, with the verdict and its sources expanded underneath. You can also read a summary of what the video actually claimed, built from the checked claims rather than from the raw transcript.
What Happens Between the Claim and the Verdict
A sentence on a page is not something a search engine can verify. It has to be turned into something checkable first.
- ●Claim extraction. The text is decomposed into discrete factual propositions. "The round was around ten billion, which was a mistake" contains one checkable claim and one opinion. Only the first goes forward.
- ●Evidence retrieval. Each claim is matched against Factagora's Temporal Knowledge Graph, which stores facts as FactBlocks: individual propositions carrying their own source, confidence, and validity window. A fact that was true in 2023 and false in 2026 is stored as both, with the dates attached.
- ●Source weighting. Evidence is graded by credibility before it is weighed, so a primary filing and an aggregator repeating it do not count the same.
- ●Verdict. The verdict, confidence score, reasoning, and the specific sources used are returned together, because separating them is what produces an unfalsifiable label.
The same pipeline is available directly through the Factagora API. DeepVerify is what it looks like when it runs in a browser instead of a backend.
Uncertain Is a Real Answer
A checker that returns a verdict for every claim is inventing some of them.
Plenty of true statements have no accessible evidence. Plenty of false ones are too recent to have been contradicted yet. When the evidence does not settle a claim, DeepVerify says uncertain and shows what it found, rather than rounding an absence of evidence up to a confident call.
This is the same failure mode that makes language models cite papers that do not exist. A model asked to produce a citation will produce one, because producing text is what it does, and nothing in the generation process distinguishes a real source from a plausible-looking one. Retrieval narrows the gap but does not close it, which is why RAG reduces hallucinations rather than curing them. Refusing to answer is not a gap in coverage. It is the part that makes the other answers worth reading.
Getting Started
Install from the Chrome Web Store, sign in with your DeepVerify account, and it runs. There is nothing to configure. Highlight a sentence and you have a verdict, with its sources, in about the time it takes to read this paragraph.
The free plan covers a few documents a day. Live audio and synced history come with Pro. Full details are on the DeepVerify page.