How to Tell if a Video Is AI-Generated
Five checks, in the order that catches the most: labels and watermarks, what the picture gets wrong, what the sound gets wrong, where it came from, and a frame check with the image detector below. No tool settles it alone, and this page says where each one fails.
Frame check. Pause the video on a clear, well-lit frame, take a screenshot, and drop it here. The detector reads that one image. It cannot hear the audio or see the motion, so use it as one signal beside the four checks below.
Drag & drop, paste, or click to upload
PNG · JPG · WEBP · HEIC — your image is analyzed privately and deleted
This is an estimate, not proof — the detector can be wrong, producing both false positives and false negatives. Don't rely on it as your only source; verify anything important. wedetect.ai is not liable for decisions made from these results. See our Terms.
1. Labels and watermarks, the fast check
YouTube, TikTok and Meta ask uploaders to label AI content and add their own labels when they detect it; look under the caption and in the video's details. Some generators mark their output: Sora adds a visible moving watermark, Google's Veo carries an invisible SynthID mark, and some tools attach Content Credentials that a checker can read from a downloaded file.
The honest limit: marks get cropped, re-encoded and removed every day, and most platforms strip file-level credentials on upload. A mark that is present is strong evidence. A mark that is absent is no evidence at all.
2. What the picture gets wrong
- Hands, fingers and teeth that change count or merge, especially when they move fast or cross something.
- Text, logos and signs that are almost legible, then change between frames. Real signage stays put.
- Edges and hair that shimmer or crawl against the background, and jewellery or glasses that flicker.
- Physics that is slightly off: liquids, cloth, smoke and shadows that do not follow the motion, objects that pass through each other, reflections that do not match.
- Continuity: background people who morph, a cup that changes shape, a pattern on a shirt that rewrites itself across a cut.
- Faces that blink too little or too evenly, mouths that lead or lag the voice, skin with no pores under a sharp lens.
Generators improve every few months, so any single tell can be absent. Several together, in a clip nobody can source, are what matter.
3. What the sound gets wrong
Listen with headphones. Synthetic speech often has no breath sounds and no room, so a voice in a kitchen sounds like a voice in a booth. Lip movement that leads or trails the words by a fraction, footsteps and object sounds that do not match what you see, and background noise that loops are all worth a second listen. A clip with the sound removed is itself a small signal.
4. Where it came from
Check the account: how old it is, what else it posts, whether it names a source. Search a screenshot of a distinctive frame in Google Lens, TinEye or Bing Visual Search to find an earlier or original copy; a clip that exists nowhere else, from a new account, with a strong emotional hook, is the standard shape of a fabrication. For anything that matters, ask for the original file rather than a forward or a re-upload; see where did this image come from for the same routine applied to stills.
5. The frame check, and what it cannot do
Our detector reads single images. A clear frame from an AI video often carries the same pixel-level signs as an AI still, so a frame check is a real signal, and it is free. Choose a frame with a face or a detailed object, not a blurred motion frame, and check two or three frames rather than one.
What it cannot do: judge the motion, the audio, or a deepfake where only the face was replaced on real footage, since that frame is mostly real. We do not offer video upload today. When we do, it will be listed here.
Frequently asked questions
Can wedetect check a video?
Not as a video, yet. What works today is a frame check: pause on a clear frame, take a screenshot, and upload it. That is a single-image check, so it says something about that frame and nothing about the sound or the motion.
Do AI videos have watermarks?
Some do. Sora and Veo add visible or invisible marks, and some tools attach Content Credentials, but marks are cropped, re-encoded or removed constantly. A missing watermark proves nothing; a present one is useful.
How can I tell if a TikTok or Reel is AI?
Look for the platform's AI label first, then check the account (age, history, whether the same clip exists elsewhere), then watch the hands, the text in the scene, reflections and the way objects behave across cuts. Sound that is slightly off from the picture is a common tell.
Are AI video detectors reliable?
Less than image detectors, and the good ones say so. Generators change every few months, and compression on social platforms removes many of the traces detectors rely on. Treat any single tool as one signal.
What are the most common tells in AI video?
Hands and fingers that change count, text and logos that do not stay legible, hair and edges that shimmer, physics that is slightly wrong (liquids, cloth, shadows), faces that do not blink normally, and background people or objects that morph between frames.
Is this free?
Yes. The frame check on this page is the same free, no-sign-up detector as the rest of the site, and the image is deleted after analysis.
Related: deepfake detector · is this person real? · AI watermarks explained · how to tell if an image is Sora