How to Tell If a Video Is AI

Short answer

To tell if a video is AI, pause it and check the details models still fumble: hands, text, physics, faces, light and sound. Then check the context: platform labels, content credentials, and where the clip first appeared. Finish with a free detector. No single check settles it, but several pointing the same way usually do.

The clip in your feed is moving, compressed and over in eight seconds. That is exactly why fakes work. The fix is boring and effective: slow it down, look at the parts a generator has to invent from scratch, and then look at everything around the clip, not just the clip itself.

Google's own help page for its verification tool lists the same starting points: garbled text, errors in hands, teeth and background details, inconsistent shadows and lighting, and repetitive patterns. The European Digital Media Observatory, a network of fact checkers, adds the context checks: who posted it, whether fact checkers already covered it, and whether the audio really belongs to the picture. Here is how to run all of it in a few minutes.

The 9 checks

1Hands and fingers

Hands are the classic giveaway because they are small, busy and constantly overlapping other things. Pause on any moment where a hand grips a cup, a phone or another hand. Count fingers. Watch for a thumb that bends the wrong way, fingers that fuse into the object they are holding, or a ring that hops from one finger to another between shots.

A clean hand proves little. Newer models get hands right far more often than older ones did. A broken hand, though, is hard to explain any other way.

2Text and signs

Real cameras record lettering as it is. Generators draw it, frame by frame, and they often lose track. Look at shop signs, number plates, jersey numbers, product labels and screens in the background. AI text tends to look plausible at a glance and fall apart when you read it: letters that are not quite letters, words that change spelling as the camera moves.

3Physics

Watch what things weigh and how they collide. Water that pours like gel, a ball that bounces too high, a door that opens without anyone touching it, a crowd where people pass through each other. Even model makers admit this is hard: OpenAI described more accurate physics as a selling point of its newer video model, which tells you older clips often missed. Physics errors are easiest to spot at 0.5x speed.

4Faces, eyes and teeth

Faces get the most attention from model builders, so the errors move to the edges. Check teeth (a single white band, or too many), earrings that change shape, glasses that melt into skin, and hairlines that shimmer. Blinking that is far too regular, or missing for a long stretch, is worth a second look. If the face belongs to a public figure saying something surprising, raise your bar.

5Light and shadow

Pick one light source, a window, the sun, a lamp, and follow it. Shadows should fall the same way for every person and object, and reflections in mirrors, windows and sunglasses should match the room. Google lists inconsistent shadows and lighting among the common artifacts. In AI clips the lighting is often flattering and even in a way real phone footage rarely is.

6Sound and lip sync

Turn the sound up. Does the voice match the room, an echo in a hall, wind outside? Do footsteps land when feet do? EDMO's guidance is direct: check the synchronization of audio and video, because fake audio can be laid over real footage. Watch mouths on hard consonants like "b", "m" and "p", where lips must close.

7Platform labels

Before you squint any harder, check whether the platform already told you. YouTube requires creators to disclose realistic AI generated or altered content, shows a label in the player or the expanded description, and may apply one automatically when a file carries C2PA metadata. TikTok said in May 2024 that it would start automatically labeling AI generated content uploaded from certain other platforms by reading Content Credentials. A label is strong evidence. A missing label is not evidence of anything, because disclosure depends on the uploader and on metadata surviving the trip.

8Content credentials and watermarks

Some generators stamp their output. OpenAI said every Sora video carries a visible watermark and embeds C2PA metadata. Google says its AI video carries an invisible SynthID watermark that a checker can read. If you have the original file, upload it to the Content Credentials Verify tool or ask Google's tools to look for SynthID. The catch: the C2PA standards body itself says metadata can be removed, inadvertently or intentionally, and a cropped repost may carry nothing. More on both in the Sora watermark explainer and SynthID explained.

9Background and continuity

Generators focus on the subject and improvise the rest. Scrub through the clip and watch the edges: a car that turns into a bush, a bystander who vanishes behind a pole and never comes out, tiles that repeat in a too-perfect grid. Google flags repetitive patterns that seem unnatural. Continuity also covers cuts: real footage from one event should agree with itself across angles.

Find the original source

Most viral fakes are reposts. Find who posted first. Take a screenshot of a clear frame and run a reverse image search; Google's help page recommends reverse image search to find original sources. Then look at the account. EDMO suggests asking whether it is anonymous, a few days old, or consistently shares the same kind of AI content. Search the claim on fact checking sites, and check whether reliable outlets report the same event. A dramatic clip of a real place that no local outlet mentions deserves suspicion.

Run a free check

Your eyes catch a lot, but they cannot read what is written inside a file. The free checker on our homepage reads the clip's file in your browser, so the video is not uploaded. It looks for Content Credentials under the C2PA specification, for AI labels that use the IPTC digital source type codes, where "trainedAlgorithmicMedia" means created using generative AI, and for the names of AI video tools in the metadata. It also shows still frames from across the clip for you to inspect, without scoring them. The result is one of four verdicts: AI generated, Likely AI, Likely real, or No AI signs found. It is free with no signup.

Read the verdict as one signal, not a ruling. "No AI signs found" means the file carried no AI label the check could read, not that the clip is proven real. EDMO says it plainly: use tools but do not rely exclusively on them. Pair the result with what you found above.

For a frame-by-frame AI score, or if the doubt is about a still image or a voice note, DeepfakeDetector.ai offers a free account with 50 scans a month and clips up to 2 minutes, and it also checks images and voice.

If the tells are what you want to get better at, the visual tells guide has annotated frames for each one. Guides on platform labels, Content Credentials and scams are grouped by topic in the AI video guides library.

Want a second opinion after the nine checks? Ask the question directly: is this video AI? The checker returns a verdict from the file's own labels, with still frames to look at. If the clip shows a real person who may have been faked, the deepfake video detector page covers that case, and if you suspect a specific generator, start with the Sora video checker notes on what a visible mark can and cannot prove.

Google's tools leave their own traces; the Veo video checker explains what to look for.

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. YouTube Help: Disclosing use of GenAI content
4. TikTok Newsroom, 9 May 2024: Partnering with our industry to advance AI transparency and literacy
5. OpenAI: Launching Sora responsibly
6. OpenAI: Sora 2 system card
7. Google DeepMind: SynthID
8. C2PA: FAQs
9. C2PA: C2PA Technical Specification 2.1
10. IPTC: Digital Source Type NewsCodes
11. Content Credentials: Verify tool

All accessed 10 November 2026.