Generative video is the fastest-moving target of all.
VideosVI
If you learned to catch fake video by waiting for the face to stop blinking, your method is a museum piece. In 2018 a well-known study showed that the deepfakes of the day barely blinked, because they were trained on still photos of open eyes. It was a genuinely useful tell for about a year. Then the people building these systems read the same paper, fed their models video instead of stills, and taught the faces to blink on schedule.
That is the pattern with every video tell. The flaw gets named, the flaw gets fixed, and the advice that worked last year quietly stops working. Generative video moves faster than any other format we cover, which makes it the hardest to give honest rules for. So let's separate what still earns its keep from what's already gone, and then get to the move that actually matters.
A single frame can be flawless while the sequence falls apart. Generative video is built frame by frame, and the model has to remember what it drew a moment ago. It often doesn't, quite. Look for:
Flicker on fine detail: jewelry, teeth, patterned fabric, text on a sign that subtly rewrites itself frame to frame.
Warping where edges meet: the hairline, the boundary between a face and the background, and especially where hands cross in front of objects. These seams are where the model's memory breaks down.
Background morph: a doorway, a parked car, or a crowd member that shifts shape, doubles, or melts when the camera pans.
The single best habit here is to watch the clip a second time and stop pixel-peeping the main subject. Stare at the corners.
When audio is grafted onto a face, the mouth and the sound fall out of step in small ways: a syllable lands a beat early, the lips form a shape the word doesn't need, the jaw moves without the right tension. It's subtle, and it's one of the more durable tells because getting mouth physics exactly right across a whole sentence is genuinely hard. Researchers build detectors specifically around mouth-region inconsistency for this reason. But "more durable" is not "reliable": on a few seconds of careful work, drift can be imperceptible.
Real light obeys the scene. When a person turns, shadows swing across the face; when they step toward a window, the highlights climb. AI video frequently keeps lighting that's painted on rather than physical: a shadow that doesn't move when the head does, a glint that stays put through a turn, skin that's lit from a source nowhere in the frame. Movement is where this cracks, so look for it during the motion, not the still moments.
The newest text-to-video systems handle gravity and collision far better than the first generation did. A dropped object now usually falls; a missed shot now usually rebounds. Usually. The failures have moved up a level: causal sequences (a cookie that's bitten but unmarked), precise object interaction, liquid behaving like a solid, a hand that passes through a railing it should grip. When something in the clip violates how the world works, even slightly, trust that more than any single frame.
Honesty is the whole point of this exercise, so here is what to stop relying on.
"It doesn't blink." Dead within a couple of years of the study that named it. Modern faces blink, and they blink at plausible rates. Blink analysis is still a real signal for software that measures rate and timing across many frames, but to the naked eye, a normal blink now means nothing.
"The hands are wrong." This was a great image tell and a decent video tell, and it has weakened sharply. Hand rendering has improved a lot on ordinary poses, and a clean hand is no longer reassuring.
"Too smooth, too perfect." Early AI faces had that waxy, poreless look. Current systems add convincing skin texture, stray hairs, and asymmetry on purpose. Smoothness is a hint at best, and plenty of real footage is over-smoothed by filters anyway.
If your detection rests entirely on the list above, you're working from an out-of-date map.
Here is the uncomfortable part. On modern, short, polished clips, human eyeballs are not dependable. A 2024 meta-analysis pooling 56 studies put overall human detection of deepfakes at roughly 55%, barely above chance, and only modestly better for video specifically (57%). Worse, people tend to be more confident than that number justifies.
There's a genuine bright spot: in at least one controlled study, humans actually beat automated detectors on deepfake video, even as those same detectors trounced humans on still images. So your instincts aren't worthless on motion. But "better than a struggling algorithm" is a low bar, and it collapses on the exact clips designed to fool you. Don't give up, but don't mistake a clean-looking video for a verified one.
When the pixels can't be trusted, switch to where the video came from. This is the part that's genuinely improved.
A cross-industry standard called C2PA ("Content Credentials") attaches a signed, tamper-evident record of how a file was made and edited. In 2026 the backing got serious: OpenAI announced it was becoming C2PA-conformant, and Google said Content Credentials verification is coming natively to Search and Chrome, with some phone camera lines embedding credentials into video at capture. A complementary approach, Google's invisible SynthID watermark, hides a signal in the content itself that's designed to survive cropping and compression; Google reports it across more than 100 billion images and videos. For more on how these labels work, and where they fail, see our deepfakes explainer.
The catch, which you should never forget: metadata can be stripped. A re-upload, a screen recording, a quick pass through a "remove watermark" service, and the provenance is gone. When Sora 2 shipped in September 2025 with both a visible badge and embedded credentials, third-party removal tools showed up within about a week, and some outputs went out missing the watermark entirely. So the absence of a credential proves nothing, while its presence is strong evidence.
Practical version: who first posted this, and would they actually have it? Does a reputable outlet carry the same footage from a second angle? Run a reverse search on a keyframe. The shared physics tells from our AI image guide apply frame by frame, and if the clip has speech, the cues in our AI voice guide catch what your eyes miss. The fastest fakes win by making you react before you check, so the real skill is being slow on purpose.
Can you reliably spot AI video just by watching it in 2026?
No. On short, polished clips, careful viewers do a little better than a coin flip, and worse than that when the clip is doing its job. Treat your eyes as one input, not a verdict.
Doesn't the 'AI faces don't blink' trick still work?
It expired years ago. The 2018 finding that deepfakes rarely blinked was true at the time; the models read that research too. Blink patterns can still help trained detection software, but a normal-looking blink tells a human almost nothing now.