When we built the claim-extraction step in verifAI, the first version made a mistake we underestimated: it tried to fact-check everything. The result was a wall of verdicts on opinions ("Spaghetti carbonara is better than amatriciana"), rhetorical questions ("Why does nobody talk about this?"), and self-evident truisms ("Water is wet"). The verdicts were technically correct and completely useless.
It turned out that a surprisingly large fraction of any short-form video is not verifiable — and the fact-checker's first job is knowing when to stay quiet. This post is the working taxonomy we ended up with. It is useful for the everyday reader who wants to be a better consumer of social video, and for anyone building a similar pipeline.
The five categories that show up in a typical reel
When you pull a 30-second reel apart sentence by sentence, what you find usually falls into one of five buckets.
1. Verifiable factual claims
These are sentences that make a specific, falsifiable assertion about the world. The hallmark is that you can in principle look up evidence for or against them.
- "Spain's unemployment rate hit 11.2% in Q1 2026."
- "The Eiffel Tower is taller than the Statue of Liberty."
- "Vitamin C does not prevent the common cold."
These are the ones worth checking. They have specific subjects, specific predicates, and an answer that doesn't depend on the reader's worldview.
2. Opinions and value judgments
A sentence with a verb like "should", "deserve", "is the best", "is amazing", "is overrated" is almost always an opinion. There is no fact to check, only a preference.
- "Carbonara should never have cream in it."
- "The new iPhone is overrated."
- "Politicians are all the same."
A good fact-checker stays out of these. A bad one ends up writing thousand-word articles concluding "depends on the criteria you use", which nobody benefits from.
3. Rhetorical and emotional framing
Most viral video is structured around emotional hooks. "Wait until you see this." "This will shock you." "Nobody is talking about this." These sentences are signalling, not asserting. They convey urgency or in-group complicity rather than information.
This bucket is harder than it looks for an automated extractor because rhetorical sentences often sit next to real claims, and the rhetorical framing can subtly shift what the claim is supposed to mean. ("Nobody is talking about this government scandal" — is "government scandal" the claim, or just an emotional label?)
4. Vague or unfalsifiable claims
These are sentences that sound substantive but evaporate when you try to write a search query for them.
- "Experts agree this is a problem." (Which experts? Which problem?)
- "Studies show that X." (Which studies?)
- "People are saying that Y." (Who?)
A claim that can't be reduced to a query the way a search engine would understand isn't a fact-checkable claim — it's a rhetorical move borrowed from the format of one.
5. Predictions and hypotheticals
A sentence about the future is, by definition, not currently verifiable.
- "This stock will double next quarter."
- "AI will replace all teachers by 2030."
- "If we keep going this way, we will collapse in 10 years."
You can fact-check the track record of predictions, or the plausibility given current evidence, but you cannot fact-check the future itself. Conflating the two is one of the most common ways for a confident fact-checker to embarrass themselves.
The two tests we use to decide
The claim extractor in verifAI applies two cheap tests before it lets a sentence through:
- Can you write a Google query for this? Not a vague query — a specific one that would return evidence one way or the other. If the only query you can write is "is X true", the sentence is not concrete enough.
- Would two careful readers, looking at the same sources, reach the same verdict? If the question is "depends on your worldview", the answer is no, and the sentence is an opinion in disguise.
Both tests are imperfect, but they catch the worst false positives. A sentence that passes both is the kind of thing you can actually defend a verdict on.
What you can do as a reader
If you want to internalise this without ever opening a fact-checking tool, the single highest-leverage habit is mentally tagging which sentence in a video is the claim. Most reels make this hard on purpose — they front-load the emotional framing, bury the actual claim in the middle, and end with a call to action. Force yourself to summarise the claim in one sentence to yourself. Often, the moment you do this, you realise that the claim is vaguer than you thought, and the video has been doing the work of making you feel certain without actually telling you what it wants you to be certain about.
Then ask the test-1 question: what would I search for to check this? If you can't come up with a query, you are probably reacting to framing, not information.
What this means for a fact-checking pipeline
For anyone building one of these systems, the lesson we learned the hard way: the precision of your claim extractor sets the ceiling for everything downstream. If your extractor lets through opinions, your synthesis model will dutifully cite sources for "carbonara should not have cream" and the user will lose trust in every other verdict on the page.
The single biggest accuracy improvement we ever made was not a smarter model. It was tightening the system prompt of the extractor with explicit rules:
- Drop sentences containing should, deserve, must, the best, amazing, ridiculous, etc.
- Drop sentences in interrogative form unless they restate a claim (rare).
- Drop sentences whose verb is in the future tense.
- Drop sentences whose subject or object is "people", "they", or "everyone" without further qualification.
Each of these rules removed a category of garbage from the verdict cards and made the remaining ones feel sharper. Not every problem in this space is a model problem — most of them are taxonomy problems.
Closing thought
The reason verifAI exists is that short-form video makes it easy to feel informed without actually being informed. Most of the work of being a better media consumer is not learning to spot lies — it is learning to spot when there isn't actually a claim being made, and reacting accordingly. That's a habit any reader can build without any tool at all.
If you have other categories you think we're missing or examples that don't fit cleanly into the five above, I'd love to hear them: fernandoruedaoliva@gmail.com.