The first reaction I see when people get an "unverified" verdict on verifAInow.es is frustration. They came for a clear answer — true or false — and they got something that sounds like a shrug. I want to explain why that reaction, while understandable, has the logic backwards.
The five verdicts, and where "unverified" sits
verifAI rates each claim as one of five things: true, false, partially_true, misleading, or unverified. The first four all share something: they require evidence. You cannot call a claim true without a source that confirms it. You cannot call it false without a source that contradicts it. You cannot call it misleading without evidence of what is being omitted.
"Unverified" is different in kind, not just in degree. It means the search turned up no evidence relevant enough to support any of the other four ratings. It is not a weak version of "false." It is a precise description of the evidential situation at the moment the check ran.
This distinction matters. If you see "unverified" and mentally translate it as "probably false," you have made an inference the data does not support. The claim might be false. It might also be true and simply underdocumented. Both are possible. The verdict is telling you: we do not know.
Why "unverified" comes up as often as it does
There are four situations that produce "unverified" verdicts reliably, and none of them are failures of the tool — they are honest acknowledgements of the limits of what is knowable right now.
Breaking news
The pipeline checks two sources: the Google Fact Check Tools API, which aggregates work from Snopes, AFP, AP, EFE Verifica, Maldita.es, and dozens of other outlets, and a live web search that fetches and reads current pages. Both lag the news cycle. A claim that entered the world in the last few hours has not been covered by any fact-checking outlet yet, and the web search will return only the original breaking story — which repeats the claim rather than verifying it.
This is not a flaw in the pipeline; it is a flaw in time. No fact-checking process, human or automated, can check something that has not yet been reported on. The honest response is "unverified, no recent sources" — not a made-up verdict stamped with false confidence.
Niche and technical claims
Some claims live in a part of the knowledge graph that is thinly documented on the public web. Specialist medicine, niche economics, obscure regulatory history — these topics exist in primary sources (journal articles, government databases, regulatory filings) that are either paywalled, not indexed, or require domain expertise to interpret correctly. The synthesis model can find a paper and summarise it, but it will not invent an interpretation when the evidence is ambiguous. "Unverified" is what you get when the model reaches the edge of what the open web can reliably settle.
Claims that are partly opinion
The claim extractor is instructed to drop pure opinions and rhetorical questions, but language is slippery. A claim like "this economic policy caused unemployment to rise" sits on the boundary between an empirical statement and an interpretive one — there may be data on unemployment, and there may be economists who disagree on causation, and no single fact-check will resolve the dispute. When a claim cannot be disentangled from contested interpretive frames, "unverified" is the most accurate rating available.
Missing primary sources
Sometimes a claim references something that should be verifiable — a study, a statistic, a specific vote — but the primary source is not publicly accessible, the study does not appear to exist, or the statistic is cited without attribution. This is related to, but different from, the niche-technical case. The issue is not that the topic is obscure; it is that the specific thing the claim points to cannot be found. I wrote about this pattern in more depth in "/A study says" is not a source.
The dangerous alternative: a tool that never says "unverified"
Here is the thing I want you to sit with for a moment. A tool that always produces one of true / false / partially_true / misleading — never "unverified" — sounds more powerful. It feels decisive. In practice, it is lying to you.
When a language model is asked to rate a claim and the evidence is insufficient, it has two options. It can say so, or it can confabulate: generate a plausible-sounding rating with a plausible-sounding explanation, composed entirely from the model's training data rather than from actual sources checked in real time. The explanation will read fluently. The sources it cites may not exist, or may not actually say what the model claims. This is not an edge case; it is one of the most consistent failure modes of LLM-based fact-checking tools.
The whole reason "unverified" exists in the verifAI schema is to give the synthesis model an honest exit. If the search returns nothing relevant, the model is not supposed to fill the gap from memory. It is supposed to say: I checked, I found nothing, I will not invent a verdict. You can read the full pipeline explanation in How verifAI fact-checks a reel — the design choice is explained at Step 6.
A model that never admits uncertainty is not a fact-checker. It is a confidence generator. Those are very different things, and the difference matters most precisely in the cases where a claim is new, niche, or contested — which is exactly when you most need reliable information.
What to do with an "unverified" verdict
The right response to "unverified" is not "it's been debunked" — that is the opposite of what the verdict says. It is also not "it must be true, otherwise verifAI would have caught it." The verdict is specifically withholding that judgment.
The practical guidance is simple: withhold sharing until you have better evidence. If a claim is important enough that people in your network would act on it, an "unverified" verdict is a reason to pause, not a green light. Find the primary source. Search for coverage by a specialist outlet. Wait for the news cycle to catch up. If the claim is trivial, perhaps it does not matter much either way.
What "unverified" is not: a debunking. Do not describe it as one. Telling someone "that claim was debunked" when the verdict is "unverified" introduces exactly the kind of misinformation the tool is trying to reduce.
The broader point
Fact-checking tools that always have an answer are optimising for the feeling of certainty, not for accuracy. The uncomfortable truth is that many claims circulating on social media cannot be definitively rated with the evidence available right now. That gap — between what people want to know and what can actually be verified — is not a problem any tool can engineer away. It is a feature of epistemics.
verifAI's job is to close as much of that gap as possible with actual evidence, and to be honest about what remains. "Unverified" is not us giving up. It is us refusing to fill genuine uncertainty with manufactured confidence.
If you want to see this in practice, there is a worked example in Case study: "No market without state" where a claim turns out to depend on contested interpretation, and the post is transparent about why even its FALSE rating is debatable.
If you have a verdict that seems wrong — especially an "unverified" that you think should have found sources — I want to hear about it. My address is fernandoruedaoliva@gmail.com.