When I built verifAInow.es I wasn't trying to replace human judgment — I was trying to automate the tedious parts so human judgment could land somewhere useful faster. The pipeline I describe in how verifAI fact-checks a reel is, at its core, the same sequence a careful person would follow by hand. The difference is speed and scale.
This post is for the moments when you don't want to paste a URL into a tool — when a friend has sent you a screenshot, or the video is embedded in a WhatsApp forward with no shareable link, or you just want to understand the reasoning rather than outsource it. Five minutes is realistic for a single clear claim. Complex or jargon-heavy claims take longer; I'll flag where.
Step 1 — Transcribe the audio and read the on-screen text
The first thing the pipeline does is pull everything the video is actually saying — spoken words and text burned into the frame — into one document. You're going to do the same thing, but by hand.
Play the video through once with the sound on. If there are captions, read them too. Then play it again and write down (or type) the specific sentence or number that made you suspicious. Don't summarise it yet — get the actual words.
This step matters because the claim is often not in the voice-over. Short-form video producers know that viewers often scroll with the sound off, so the bold assertion ("Scientists confirm…", "Government admits…") is usually the title card or the text overlay, not the narration. If you only half-listen, you can miss where the claim actually is.
Step 2 — Isolate the single checkable claim
A two-minute reel can contain a dozen statements. Most of them are opinions, metaphors, jokes, or rhetorical questions — none of those are checkable in any useful sense. Your job is to find the one or two sentences that are:
- Specific enough to be true or false (a number, a named event, a named person's action)
- Stated as fact, not as "I think" or "many people say"
- Important enough that it would actually matter if it turned out to be wrong
The pipeline uses the following rules when it strips the transcript: drop opinions, drop jokes, drop rhetorical questions. Hold each remaining statement against this test: If I could prove this is false, would the video's core argument fall apart? If yes, that's your claim.
Write it out as a single, self-contained sentence. If understanding the claim requires context from earlier in the video, fold that context in. For example: not "He was convicted of fraud" but "Former health minister Juan García was convicted of fraud in 2023."
Step 3 — Search for the factual core, not the framing
Now take your claim and strip out the emotional language and the video's framing before you search. "Shocking: scientists PROVE fluoride lowers IQ in children" becomes fluoride IQ children study as a search query. "The government is hiding cancer cure research" becomes cancer cure research funding government.
Run that query, but add one more term that forces the results toward primary sources. Good terms to append:
site:who.int,site:.gov,site:.edu— for health, law, sciencepubmed,doi,clinical trial— for medical claimsofficial register,ruling,judgment— for legal claims- The name of the relevant institution, ministry, or journal
The goal is to get to the document the video is either accurately reporting on or distorting. Most viral misinformation is not pure invention — it is a real study, a real event, or a real statistic that has been stripped of its caveats and repackaged with a stronger conclusion than the evidence supports. Finding the source document lets you compare what the source actually says against what the video claims it says.
If Google returns nothing solid in the first two pages, Bing or DuckDuckGo sometimes surface different results. If you are checking a claim about Spain or Latin America, searching in Spanish often turns up completely different sources — and sometimes the claim simply does not exist in Spanish-language media at all, which is itself a signal.
Step 4 — Check a dedicated fact-check database
Web search surfaces everything. Fact-check databases surface only the work of professional verification teams, which is more useful when a claim is viral enough to have been checked before.
These are the ones I use and that the pipeline queries:
- Google Fact Check Tools (factchecktools.google.com) — aggregates tagged fact-checks from hundreds of outlets worldwide. Type the key phrase from your claim into the search box. If it has been checked by AP, Reuters, AFP, Snopes, or any partner outlet, it appears here with the original verdict and a link.
- Maldita.es — the best Spanish-language operation. They cover Spain, and increasingly Latin America, with original reporting rather than just aggregating others. Strong on health claims and political manipulation.
- EFE Verifica — Spain's national news agency's verification desk. Especially reliable for political and institutional claims about Spain.
- Snopes — the oldest English-language fact-check operation, strongest on US politics and viral internet content.
- AP Fact Check and AFP Fact Check — wire agencies' verification teams; good coverage of international claims.
Search your claim phrase in each. If you find a match, read the full article, not just the verdict label. Fact-checkers explain their reasoning and cite their sources; that reasoning is the useful part. A "false" label without evidence is no better than the original video.
If no database has checked the claim, that absence is not evidence the claim is true. It might mean it hasn't circulated widely enough to attract professional attention yet. Keep going.
Step 5 — Decide, and be willing to land on "unverified"
By now you either have evidence or you don't. Here's how to read the situation:
True: You found a primary source — the document, the official record, the peer-reviewed paper — that says exactly what the video claims, with the same scope and caveats. The claim is accurate.
False: You found a primary source that directly contradicts the claim, or a credible fact-checker who traced the claim to a misquote, a manipulated image, or a document that says something else entirely.
Partially true: The source exists but the claim has been inflated, stripped of a key qualifier, or taken out of context in a way that changes its meaning. This is the most common case for viral content that starts from real information.
Misleading: The factual content is technically accurate but the framing, selection, or juxtaposition is designed to create a false impression. A video that shows a real statistic from 2008 and presents it as current is misleading, not false.
Unverified: You searched in good faith for 5 to 10 minutes, found nothing solid either way, and the claim is not outrageous enough to dismiss on priors alone. This is a real, legitimate verdict. It is not a failure. It means the evidence is absent, and absent evidence is different from evidence of absence.
The pipeline is built around this same five-option scale, and "unverified" exists specifically so the tool — and you — have an honest out rather than manufacturing confidence that isn't there. I wrote more about what that verdict actually means in what "unverified" really means.
A note on speed vs. depth
Five minutes is enough to dispose of most clearly false viral claims, because most of them have been checked already and the Google Fact Check Tools result comes up in the first search. For a fresh claim, or one involving specialist knowledge — medicine, finance, law — give yourself more time and lower your confidence threshold. The hardest honest sentence in verification is "I don't know yet."
How this maps to what verifAI does
The pipeline on verifAInow.es runs these same five steps, but in parallel and in seconds rather than minutes: it transcribes audio and reads on-screen text simultaneously, splits the transcript into atomic claims, and then fires the Google Fact Check query and the web search at the same time for each claim, landing on the same five-verdict scale. The advantage is speed and the ability to check every claim in a reel, not just the one that caught your eye.
The disadvantage is that an automated tool can miss context that a human reader picks up immediately — a sarcastic tone of voice, an obvious parody format, a niche reference that makes the claim's meaning clear in context. That's why I want the manual version documented alongside the automated one. The two approaches are most useful together.
If you want to see the automated version run, paste any reel or TikTok link into verifAInow.es and watch the pipeline panel. If you catch a verdict that looks wrong, the transcript and source links are all there — you can trace every step.