How to Spot AI Video

Short answer

To spot AI video, pause and step through the clip looking for errors a camera cannot make: hands with the wrong fingers, text that changes as you read it, objects that ignore weight, faces that shimmer at the edges, shadows that disagree, and sound that does not fit the room. Two or more tells together are a strong signal.

A camera records what is in front of it. A video generator invents every frame, and it is best at the things it has seen most and worst at small, fiddly, rule-bound details. That gap is where the tells live. Google's verification help page names the common ones: garbled text, errors in hands, teeth or background elements, inconsistent shadows and lighting, and unnatural repetition. Fact checkers in the European Digital Media Observatory network add out of focus contours, lip sync and audio that does not match the picture.

Two habits make every tell easier to see. Watch at half speed, and pause on moments of contact: hands touching things, people crossing paths, objects landing. Generators do well with a single subject standing still. They struggle when things interact.

Bodies and faces

Hands are still the first place to look. Count the fingers when a hand wraps around a cup, a steering wheel or another hand. Look for a thumb on the wrong side, knuckles that bend backwards, or fingers that sink into the object instead of resting on it. Jewellery helps: a ring or watch that changes position between frames is a giveaway.

Faces get the most training attention, so errors drift outward. Check teeth (a smooth white band instead of separate teeth), ears that change shape when the head turns, earrings that appear and disappear, and glasses whose frames melt into the cheek. Hair is hard to generate: watch the hairline against the background for a soft shimmer, the kind EDMO calls out of focus contours. Blinking that runs like a metronome, or a long stare with no blink at all, deserves a second look.

Bodies in motion give themselves away at the joints. Watch knees on stairs, elbows during a wave, and feet on the ground. Feet that slide slightly while standing still, a habit sometimes called foot skating, are common in synthetic people.

Text and signs

Lettering is a reliable tell because a real camera simply copies it. A generator has to draw it again on every frame, so text drifts. Pause on street signs, shop fronts, jersey numbers, book spines, product labels and phone screens. AI text often looks right from a distance and fails up close: letters that are not quite letters, a word that changes spelling as the camera pans, numbers that reshuffle. Google lists inconsistent or garbled text as a top artifact.

One caution: real footage can have blurry or low resolution text too. The tell is text that changes, not text that is hard to read.

Motion and physics

Weight is hard to fake. Watch what happens when something falls, pours, collides or bends. Common errors include water that moves like syrup, a glass that shatters before it hits the floor, hair and clothing that do not react to wind or movement, and people or cars that pass through each other for a frame.

Model makers know this is a weak spot. OpenAI presented more accurate physics and synchronized audio as improvements in its newer video model. Read that the other way round: physics was hard enough to be worth announcing, and older clips that still circulate often get it wrong. Newer clips get it wrong less often, so treat a clean physics read as neutral, not as proof.

Camera movement matters too. Real handheld footage shakes in a messy, uneven way. Many AI clips glide like a drone even when the scene pretends to be a phone video.

Light and shadow

Find the light source and follow it. In a real scene every shadow agrees on direction, softness and length. In AI clips you will sometimes see two people casting shadows at different angles under one sun, or a person with no shadow at all on a bright pavement. Reflections are another weak point: check mirrors, shop windows, sunglasses and puddles. A reflection that shows a different pose, a missing person, or a room that does not match is a strong tell. Google names inconsistent shadows and lighting among the common errors.

AI lighting also tends to be too pretty. Golden glow on every face, perfect rim light, no harsh overhead fluorescents. That alone proves nothing, but it fits a pattern.

Sound

Turn the volume up and close your eyes for a few seconds. Real recordings carry room tone: echo, traffic, wind, the hum of a fridge. Synthetic audio is often too clean, or the background never changes even as the camera moves outdoors. Then watch mouths. Lips should close on "b", "m" and "p". EDMO recommends you check the synchronization of audio and video, because a fake voice can be laid over genuine footage. In that case the picture passes every visual test and only the sound is wrong.

When your eyes are not enough

The best AI clips now pass a casual look, and platform compression smears exactly the details you need. That is the point to stop squinting and gather other evidence.

Check for a label first. YouTube requires creators to disclose realistic AI generated or altered content and may label clips automatically when they carry C2PA metadata. Some generators also mark their output: Google says its AI video carries an invisible SynthID watermark, and OpenAI said Sora videos carried a visible watermark plus C2PA metadata. The Sora watermark explainer covers what a visible mark does and does not prove.

Then find where the clip came from. The nine-check method walks through source and context checks alongside the visual ones.

Finally, run a check. The free AI video checker reads the clip's file in your browser for the records some AI tools write into their files: Content Credentials under the C2PA specification and AI labels that use the IPTC digital source type codes. It also shows still frames from across the clip, so you can go straight to the moments worth pausing on. You get one of four verdicts: AI generated, Likely AI, Likely real, or No AI signs found. It is free with no signup. "No AI signs found" is not proof the clip is real; many AI clips carry no label, and it is one signal among several. EDMO's advice holds for every tool: use them, but do not rely on them exclusively.

If you want a frame-by-frame AI score, or you also want to check a still image or a voice message, DeepfakeDetector.ai has a free account with 50 scans a month for clips up to 2 minutes, and it checks images and voice too.

You can also start from the homepage detector and come back here with its frames. For labels, watermarks and the law, browse every guide by topic.

Spotted one of these tells in a clip of a real person? The free deepfake video check walks through that case. For clips you found on a feed, the YouTube AI video checker explains the platform's own labels. Written and kept current by Tessa Ostrowski.

Sources

1. Google, Gemini Apps Help: Verify AI-generated images, videos, and audio
2. European Digital Media Observatory: Tips for users to detect AI-generated content (archived copy of 14 May 2026)
3. OpenAI: Sora 2 system card
4. OpenAI: Launching Sora responsibly
5. YouTube Help: Disclosing use of GenAI content
6. Google DeepMind: SynthID
7. C2PA: C2PA Technical Specification 2.1
8. IPTC: Digital Source Type NewsCodes

All accessed 10 November 2026.