When a video goes viral with an alarming caption, there are two very different things it could be: real footage from an earlier event that someone has relabelled as today's news, or footage that was never filmed at all — generated frame by frame by a model. These two problems get lumped together under "fake video," but they require completely different methods to investigate. Confusing them wastes time and leads to wrong conclusions.
This post is about the manual side of video verification. I want to be upfront about one thing before we get into it: verifAInow.es checks the claims in a video against sources — it does not tell you whether the footage itself is authentic or AI-generated. That boundary is deliberate. Checking claim truth and checking footage provenance are different jobs, and trying to bolt them together poorly is worse than being honest about where each tool stops. What follows is the manual skill that complements the automated claim check, not a description of something verifAI does.
Problem one: old footage passed off as new
This is by far the more common manipulation. It does not require any technical skill to execute — anyone with a download-and-repost workflow can do it. A clip from a protest in 2019 gets shared with a caption saying it is happening now. A weather event from one country gets credited to another. The video itself is completely genuine; the lie is in the framing.
Reverse image search on keyframes
The first thing I do is pause the video at a frame with a clear, distinctive image and take a screenshot. Then I drag that screenshot into Google Images, TinEye, or Yandex Images. Yandex is worth trying even if you default to Google — it indexes more visual content and often surfaces older matches. If the clip has been circulating for a while, this frequently turns up the original upload with its original date and caption.
The choice of keyframe matters. The opening shot is often the most generic (a crowd, a sky, a street corner). A better choice is a frame with distinctive details: a specific building, an unusual sign, a crowd configuration that looks unusual, a timestamp burned into the footage. The more specific the frame, the more likely a reverse search will surface an exact match.
Cross-checking against the claimed date and location
If reverse search comes up empty, I look at what the video itself contains. Seasonal cues: the angle of light, what people are wearing, the state of foliage. A summer event should not show bare deciduous trees; a coastal city should not have mountains in every wide shot.
Signage is an underused source. Street signs, licence plate formats, police uniform markings — all jurisdiction-specific and often date-specific. A sign in a script that doesn't match the claimed country is a red flag worth following.
Weather services like Weather Underground and Ventusky archive historical conditions for most cities. A clip claiming to show a hurricane in a city that had clear skies that day is worth investigating. Once I have a candidate original upload, I check the platform's own record: YouTube shows the upload date, and the Wayback Machine at archive.org captures many videos before they go viral in a new context.
Finding the original context
The goal is not just to show that the video is old — it is to establish what it actually showed when it was originally shared. A clip of a protest can be true reporting from 2019 and still be misleading when reused in 2026 to claim something different is happening today. The original context is part of the fact.
International fact-checking organisations — Snopes, AFP Fact Check, Maldita.es, EFE Verifica — maintain searchable archives and have often already done this work. Searching the most distinctive detail of the footage (a place name, a visible sign, the alleged date) in those archives before starting a fresh investigation saves time.
Problem two: AI-generated footage
Synthetic video has improved fast enough that any technique that worked two years ago is already less reliable. I want to be honest about that before listing what to look for, because a checklist of artifacts can create false confidence.
The artifacts that still give it away — sometimes
Hands remain the most consistent weak point in current generative video. Fingers merge, split, gain extra joints, or pass through objects. The artifact is often brief — a fraction of a second — and it helps to scrub through the clip frame by frame rather than watching at normal speed.
On-screen text is another tell. Generative models struggle to render legible, stable text. Letters shift shape between frames, words contain impossible character combinations, and text that should be static often wobbles slightly. This applies both to text that is meant to appear in the scene (a newspaper, a sign) and to titles added as part of the generation prompt.
Physics breaks down in specific ways: liquids move without inertia, cloth dynamics feel wrong, fire and smoke have a looping or procedural quality rather than the chaotic specificity of real combustion. Objects that should have mass — a falling tree, a crowd surging — sometimes move as though they are weightless.
Faces in crowd shots are often a weak point even when the primary subject looks convincing. Background faces can have the blurred, symmetrical quality of generated faces rather than the variety of real human faces. The eyes of background figures sometimes track nowhere, or have a glassy uniformity.
Motion blur is handled inconsistently. Real cameras apply motion blur physically — fast movement blurs in proportion to shutter speed and the direction of movement. Generated video sometimes applies blur globally, or misses it on objects that are moving fast enough that they should be blurred.
Why these checks are getting less reliable
Every artifact I just listed has been partially addressed in successive model generations. The hands in video generated today are better than the hands from two years ago. Text rendering has improved. Physics models are being incorporated into generation pipelines. There is no reason to expect this trend to stop.
This matters for how you use these checks. Spotting an artifact is useful evidence, but failing to spot one is not clean evidence the video is real. The absence of obvious artifacts in a specific clip tells you less and less each year.
The more durable method is provenance: can you find the primary source? Was this footage published anywhere before the context it is now appearing in? Does any credible news outlet report an event that would explain this footage? AI-generated video almost never has a paper trail — it has no camera, no location, no witnesses. Real footage, even footage of something unusual, typically has at least one other source.
Metadata and source checks
Tools like InVID / WeVerify can sometimes recover or infer EXIF metadata from embedded camera fingerprints — GPS coordinates, device model, timestamp — even after a platform has stripped the obvious fields. Metadata that contradicts the claimed location or date is useful evidence. The account posting the video is also part of the check: a newly created account with a high ratio of outrage content and no engagement with local issues that would be natural for someone actually present shifts the probability, even if it proves nothing on its own.
Where verifAI fits
If the video contains factual claims — a number, a statistic, a named person saying they did or did not do something — you can paste it into verifAInow.es and get those claims checked against Google's fact-check index and live web sources. That is a different job from footage verification: not "is this footage real" but "is what this person is saying true." A synthetic video making a false claim about a named politician is both problems at once. Use the manual techniques above for the footage; use the automated pipeline for the claims.
The honest limitation
None of this is a definitive test. Reverse image search misses images that have been mirrored, colour-shifted, or cropped. AI detectors — I have not listed any specific tools deliberately, because the state of that space changes faster than any blog post can — have false positive and false negative rates that vary significantly by content type.
What these techniques do is shift probabilities and identify specific lines of inquiry. For social media verification that is often enough to reach a conclusion. When it is not, the honest answer is "unverified" — which is the same answer verifAI gives when the claim evidence is thin. That is not a failure mode; it is the only responsible output when the evidence does not support a stronger one.
For more on how to read a verdict and what the source links actually mean, see Five signs a reel is misinformation and Fact-check a video yourself in five minutes.