A video lands in your feed. A minister says something outrageous, or a soldier surrenders, or a CEO confesses. Within twenty minutes it's on three group chats and someone has already screenshotted the worst frame. The question you should ask is not "is this shocking?" but "who rendered these pixels?"
Spotting deepfake videos in news footage is a skill, not a talent. It has rules, shortcuts, and a few traps that catch even people who do this for a living. Some of what follows comes from my own work checking footage, some from the errors I made getting there.
Key takeaways
- Slow the video down. Most artifacts live in a handful of frames.
- Audio is still the weakest link in almost every deepfake — listen for the breath.
- Provenance beats perception: check where the file came from before you trust what you see.
- Human eyes alone catch a coin flip's worth of fakes. Treat your instincts as a starting point, not proof.
- The three-finger test is pop-science, not evidence.
- A deepfake in a news broadcast usually fails at the source level, not the pixel level.
Why news footage is the hardest place to catch a deepfake
Here's the thing about news video: it's already low quality. Broadcast compression, mobile uploads, grainy phone interviews shot through a car window. A face in a news clip often has fewer usable pixels than a deepfake needs to betray itself. Compression smooths over the exact artifacts you're looking for.
Which means the standard advice — "count the teeth, check the ear shape" — is weaker here than on a clean studio clip. In a controlled interview with a locked-off camera, a fake tends to stand out. In a shaky street recording forwarded through four apps, it hides.
I learned this the hard way. A clip of a local politician at a rally was circulating in a regional group. I froze a frame, zoomed in on the jawline, and convinced myself the edge was wrong. I was wrong. It was real footage, heavily compressed, shot at a bad angle. Saying so publicly would have been embarrassing. Compression artifacts look a lot like generation artifacts, and that confusion cuts both ways.
What actually distinguishes news context
News footage has something a random viral video often doesn't: a paper trail. There's a camera operator, a producer, an edit timeline, a broadcast log, a wire service. That trail is where fakes most reliably break — not in the pixels.
How can I identify deepfakes?
You identify a deepfake by cross-checking three layers in order: the source, the file, then the picture. If the source is solid and verifiable, you can stop at layer one. If it isn't, drop down a layer.
Layer one — the source
Where exactly did this video come from? Not "someone posted it" but the original upload. Who posted first? Does that account have a history? Is the clip available on the official channel of the person or organisation shown in it? If a minister allegedly said something on camera, the same footage should exist on a broadcaster's own feed or an agency's wire.
A quick search for a longer version is often decisive. Deepfakes are usually short — twenty to forty seconds — because longer clips demand more compute and more consistency. When the "full video" doesn't exist anywhere, that's a signal.
Layer two — the file
Files carry fingerprints. Most platforms strip metadata when you upload, so this only works if you can get the original file. But when you can, check:
- Creation date and device information — a clip supposedly shot on a phone will name a phone.
- Whether the file is a re-encode of something older. Reverse-image and reverse-video searches help here.
- Content credentials, where a publisher attaches them (more on that below).
Layer three — the picture and sound
Only now do you look at the face. Slow the video to a quarter speed and watch for the following:
- Lip-sync drift that gets worse over the length of the clip — often the clearest tell in a talking-head video.
- Blinking that's too regular or too rare. Real people blink unevenly, especially under stress.
- Edges that shimmer — hair against a busy background, glasses frames, the boundary of the jaw.
- Inconsistent reflections in the eyes, or eyes that don't move quite together.
- Teeth and tongue rendered as a vague blur. Interior of the mouth is still a weak spot in generated faces.
- Light direction that shifts mid-clip without anyone moving a lamp.
None of these is proof on its own. Two or three together, in the same stretch of footage, is worth taking seriously.
What is the 3 finger test for AI?
The three-finger test is a folk method — show a subject three fingers and see whether the generated hand renders them correctly. It gained traction because early image and video generators were notoriously bad at hands, and it's a fast, cheap check you can do in a video call or a livestream.
Treat it as a party trick, not a diagnosis. Hand rendering has improved dramatically; a modern model handles three fingers fine. Meanwhile a real person might have an injured hand, or hold up three fingers at an angle the camera flattens. I've seen people "debunk" genuine footage with this test and feel very pleased about it.
What to do instead in a live call
If you suspect a live video call is synthetic, ask for something unpredictable and ask for it twice. Turn your head and ask the other person to describe what changed. Ask them to hold a sheet of paper over half their face. Cheap generation pipelines struggle with novel, unscripted interactions because they can't precompute them.
Is it illegal to view deepfake videos?
In most places, watching a deepfake is not a crime, though the rules vary by country and are shifting. What is usually illegal is creating or distributing one for fraudulent purposes, or using someone's likeness without consent — especially sexual imagery or content built to defraud. Viewing it in order to report on it, fact-check it, or warn others is very different from spreading it.
Here's the practical point. The act of sharing is where a viewer crosses from bystander to distributor. Forwarding a deepfake "to warn people" without context does the fake's work for it. I've stopped doing that entirely — I quote the claim, describe the clip, and link to a debunk instead of the video itself.
If you find a deepfake of yourself
It happens more than people realise, and mostly to people with a public presence. The sequence that works: capture the URL and a timestamped screenshot, report it to the platform using its impersonation route, then contact the original publisher if there is one. Preserve everything before you report — some platforms remove content only after you've filed, and you'll want the evidence.
How can I check if a video is AI-generated?
Use the tools as a second opinion, never as a verdict. There are free online detectors that analyse a clip and return a probability, and for a quick triage they're useful. Their error rates are real, though — particularly on compressed news footage, which is exactly the input you're most likely to have.
The more robust layer in 2026 is provenance. Content credentials — cryptographic signatures attached at the point of capture by some cameras and phones — let you check whether a file was recorded by a real device and whether it's been altered since. Adoption is patchy, but news organisations increasingly sign their output, which is a genuine advantage when you're verifying a broadcaster's clip rather than a stranger's upload.
| Method | Speed | Reliability on news footage | Best use |
|---|---|---|---|
| Visual/audio inspection (slowed down) | Minutes | Moderate — compression masks artifacts | First pass on a suspicious clip |
| Source and origin check | 5–30 minutes | High when a paper trail exists | Any clip tied to a named person or outlet |
| Free online detectors | Seconds | Low to moderate | Triage, never a final call |
| Content credentials | Seconds | High where present | Checking a signed publisher file |
| Contacting the original outlet | Hours | Very high | Anything you intend to publish |
What a real case looks like
The Zelensky deepfake that appeared on a Ukrainian news channel's own website is the case I return to. It was caught within hours — not because viewers spotted a bad jawline, but because the source was checkable and the person depicted could deny it publicly. The fake was crude by today's standards. The response was fast because institutions moved, not because eyes were sharp.
The newsroom habit worth copying
When a broadcaster verifies footage, the process is usually boring and unglamorous: two people look at it separately, one of them calls the originator, and nobody publishes until the chain from camera to screen is accounted for. You can copy a stripped-down version.
- Before you react, find the earliest instance.
- Before you share, find a second outlet covering the same clip.
- Before you believe, check whether the person shown has said anything about it.
I'll be honest about my own conversion: I used to skip the first step because it felt slow. It isn't. The original upload usually settles the question in under two minutes, which is faster than squinting at a mouth for half an hour.
Where this leaves you
The uncomfortable truth is that a determined faker with decent resources will beat your eye. The good news is that most deepfakes circulating as news aren't made by determined people with decent resources — they're rushed, short, and posted by accounts with no history. That's the gap you exploit.
So the question I'd leave you with isn't "did this look fake?" It's the one that's harder to answer and far more useful: if this video turned out to be false, would anyone be able to prove it in an hour? If the answer is no — no original file, no named source, no one willing to stand behind it — you already know how much weight to give it.
