Home › How to check if a video is a deepfake
How to check if a video is a deepfake
Nearly every guide tells you to squint at the pixels. That advice is a decade out of date, and it fails on exactly the videos that matter.
Fact-Check · last reviewed 2026-08-24
The short answerTo check whether a video is a deepfake, do not start with the pixels. Find the earliest upload, establish when the event happened, and test the claim attached to it against a primary source. Detectors that score a file as fake or real are unreliable in both directions, so treat any score as a hint and never as evidence.
Why detectors are the wrong first step
A detector answers a question no one should stake anything on: does this file statistically resemble the output of a generator. It cannot tell you whether the event happened. It fails on re-encoded footage, which means almost everything that travelled through a social platform, and it fails differently for every generator released after it was trained.
We measured that pattern in a neighbouring field and published it. Across 240 texts, half written by humans before ChatGPT existed and half generated in three different styles, detection was not reliable enough to put in front of a user. Detectors are also documented to falsely accuse people writing in a second language. That is why we deliberately do not ship an AI-text detector or an AI-image classifier, and why any product that sells you one as proof is selling confidence rather than evidence.
A wrong accusation is not a smaller error than a missed fake. It is the same error pointed at an innocent person.
The four checks that actually decide
- Find the earliest uploadReverse image search a distinct frame. The oldest copy carries the real context, and a genuine event usually has several independent uploads within hours.
- Establish the date of the event, not the postTake the most specific thing in frame, a building, a livery, a uniform, a shopfront, and search that with the event word. If coverage is two years old, so is your video, however real it looks.
- Test the claim, not the fileA deepfake and a real clip with a false caption do identical damage. Check what the video is being used to assert, against a primary document: a court record, a statement, a dataset.
- Ask who benefits from you sharing itNot a mood question. If a clip arrives precisely when it is most useful to someone, that timing is evidence about the clip's purpose, and it tells you how hard to look.
What holds up and what does not
| Approach | What it establishes | How it fails |
|---|
| Detector score on the file | Statistical resemblance to generated output | Re-encoding, new generators, and false accusations in both directions |
| Reverse image search on a keyframe | Earliest known appearance and original context | Fails if the clip is genuinely new |
| Event date versus post date | Whether the implied now is real | Needs one identifiable detail in frame |
| Primary source on the claim | Whether the assertion is true at all | Slower, and some documents are paywalled |
| Provenance metadata (C2PA) | Signed capture and edit history | Only present if the whole chain supported it, and stripped by most platforms |
The bottom four survive contact with real footage. The top one is the one most tools sell.
Where a tool helps, and where it cannot
What can be automated is the boring, decisive part: gathering sources, comparing the event date with the post date, and checking that a quoted sentence really appears in the document it is credited to. That is what wyper Fact-Check does, and every statement in the result is labelled either with a quote located verbatim in the source or as unsourced.
What cannot be automated is intent. A newsroom reposting archive footage with context and an account manufacturing outrage are indistinguishable to any model. We report the time gap as its own finding and leave the judgment where it belongs. On Full Spectrum, a Triple-AI reading puts three models on one shared evidence set, so a disagreement is a different reading of the same material rather than a different search.
Questions
Can you tell a deepfake by looking at the hands or the eyes?
Sometimes, and less every month. Artefacts in hands, teeth, jewellery and reflections were reliable in 2023 and are now largely gone in current generators. Worse, compression produces similar artefacts in genuine footage, so the tell cuts both ways. Treat visual oddities as a reason to check the origin, never as a verdict.
Are deepfake detectors accurate?
Not accurately enough to accuse anyone. They degrade on re-encoded video, which covers almost everything reposted on a platform, and they fail on generators newer than their training data. We measured the same effect for AI-written text across 240 texts and chose not to ship detection at all.
What is the fastest single check?
Compare the date of the event with the date of the post. It takes about ten seconds, needs no tools, and catches the most common video hoax, which is real footage presented with a false now.
Does wyper detect deepfakes?
No, and we say so plainly. It verifies the claim attached to the video: it gathers sources, checks whether quoted sentences appear in them, and reports the gap between the event and the post. It does not run a classifier over the pixels, because that is the part that does not work.
What about audio deepfakes and cloned voices?
Harder, because audio carries fewer contextual clues than video. The same order applies: find the earliest copy, check whether the claimed speaker was where and when the recording says, and look for an independent record of the statement. A cloned voice saying something the person never said elsewhere is the pattern to expect.
Check a claim and look at the evidence yourself
Free: 5 checks a day, no account needed. Every statement tells you whether a quote backs it.
Read on
Ask any fact-checker for their error rateFour claims, four published measurements, including the numbers that count against us.
How to check if an image is AI-generatedWhy we refuse to ship a classifier, and what works instead.
wyper