I find a lot of people seem to think they can glance at an online video and quickly figure out if it’s real or not.
They can’t.
Studies put human detection rates near chance for good video “deepfakes” and AI generated content.
The technology has gotten so advanced that there’s no easy, reliable way to spot every deepfake or AI-generated video by eye alone. Discernment is important because a convincing clip can drive outrage, move markets, or worse. In 2024 there was a case involving a deepfaked video-call and participants made roughly $25 million in financial transfers based on the call. Several years ago on my TV program “Full Measure,” I spoke with experts on national security fears that a good deepfake of a president making a statement could start a war before the video was ever exposed as a fraud.
Even in today’s environment where people are on alert for fakes, videos can feel more “real” than text, so they spread faster and hit harder. Knowing the common tells—and, more importantly, how to check context and source—gives you a better chance of pausing before you share, act, or believe.
Read on for details.
Here’s a practical guide for detecting AI videos.
Length
One of the most consistent patterns I’ve noticed is that many realistic AI videos and deepfakes circulating on social media are short: often 6–10 seconds. Digital forensics expert Hany Farid has pointed out that the vast majority of clips he’s asked to verify fall in that range. The fact is: generating longer, coherent video is still computationally expensive for many tools, and longer clips give the AI more chances to glitch.
Missing context
These clips frequently show only the dramatic moment: the confrontation, the shocking statement, the heartwarming rescue, or the “gotcha.” They skip what led up to it and cut off before the aftermath.
Framing
Short, tightly framed emotional clips that leave you with almost no surrounding information are a strong reason to consider they may be fake.
Editing
Creators also stitch short segments together, which can produce noticeable cuts every few seconds.
Visual and audio clues
We’ve already learned to look for the obvious signs of extra fingers and limbs. I still think it’s weird that AI can’t always figure out humans generally have four fingers and a thumb. Even when it’s instructed to create that kind of a hand, it sometimes cannot.
I’ve been watching Medicare commercials that clearly use AI-generated people and images. The lips don’t perfectly match the sounds coming from them. The voice seems too perfect.
But what about less obvious clues and better deepfakes?
Slow the video (0.25x or 0.5x if possible) and pause on movement. Look for lip-sync and mouth issues. Real lips usually close fully on plosive sounds like “p,” “b,” and “m.” Deepfakes and lip-synced AI often drift by a frame or two, or the mouth interior looks off.
Blinks that are too regular, too rare, or mistimed with speech are another possible sign. So are eyes that look glassy or have mismatched reflections and expressions that don’t quite match the words or the rest of the face.
Watch for warping, smearing, or a soft “halo” around the jaw, hairline, ears, or where the face meets the background—especially when the head turns quickly or something crosses in front of the face such as a hand or a microphone.
Unnaturally smooth or waxy skin compared with the neck or hands can be a sign of deepfakes. So are teeth that look like a solid white block.
Be on the lookout for shadows pointing the wrong way, lighting on the face that doesn’t match the room, objects that float or pass through each other, or garbled text on signs, screens, or clothing. Sometimes background elements morph or behave impossibly.
Voices that sound too clean or flat are suspect. So are missing natural room tone— the sort of natural background sounds in a room— breathing, and pauses. At times you may detect the emotion doesn’t match the delivery.
These clues are more reliable indicators of AI videos when you find several of them together.
