AI video has gotten good enough that your scroll-past glance won't catch it anymore. The fakes still leave evidence, though. You just have to know where it hides — and build a thirty-second habit of looking. Think of it the way you think of phishing emails: ten years ago we all learned to check the sender before clicking. Same skill, new medium.
Ask the boring questions first
Before you study a single pixel, interrogate the post. Who uploaded this? Where did it show up first? If a supposedly major moment exists only as a short, cropped clip with zero context, that's already a red flag. Where's the full video? Who else has it? Real events produce multiple angles, longer cuts, and follow-up reporting. Fakes tend to live as one orphaned clip that nobody can trace.
Check the edges, not the eyes
When you do look closely, look in the right places. Pause and step through frame by frame — the artifacts cluster at boundaries: jawline, hairline, ears. There's often a faint mask effect, where skin tone or texture changes abruptly right at the edge of the face.
Then watch for flicker. Generated video is bad at keeping details consistent between frames, so things mutate: teeth come and go, skin texture shifts, a mole drifts or disappears entirely. Lighting slips too — a face lit inconsistently as it moves, or the subject's shadow pointing one way while everything nearby points another.
Mute it
Here's an underrated trick: kill the sound and watch the lips. Audio carries you past sync problems — once it's gone, mismatched mouth movements become obvious fast.
What about detection tools?
They've gotten better, and they're worth using when it matters. Current systems analyze facial movement, lighting, voice sync, and editing patterns to estimate whether footage was manipulated. The cleverest method going reads something AI doesn't fake: your heartbeat. Real blood flow causes tiny rhythmic color changes in skin, and systems like Intel's FakeCatcher pick that up through remote photoplethysmography.
One caveat every serious guide repeats: these tools give estimates, not verdicts. A detector score is one input. It's not the answer.
Stack your checks
No single test is decisive — that's fine, because you're not running one test. Verify the source. Step through frames. Watch the edges. Mute and rewatch. Run a detector if the stakes are real. Then see if anyone trustworthy has confirmed the thing actually happened. Most fakes fail somewhere in that stack.
And the last line of defense isn't software at all. It's the pause before you hit share. A fake that stops in your feed is a fake that lost, whether or not you ever proved it frame by frame.
Image: freestocks.org, via Pexels
