If you want to know how to tell if a video is AI-generated, or SI-generated as we say on this site, start by lowering your expectations of your own eyes. Video generators now produce clips with natural light, believable camera shake and synchronised sound, and controlled studies show that people guess barely better than a coin toss on the best of them. The good news is that a video carries more evidence than a single frame: labels, watermarks, file credentials, motion over time and, above all, a history you can trace. This guide walks through those checks in the order that saves the most time, and shows what happened when we ran three clips through our own video detector.
The short version. Look for a platform label or a visible watermark first. Then check the file for Content Credentials or a SynthID watermark if you have the original. Watch the clip several times with your attention on motion and physics, not on how sharp it looks. Finally, trace the first upload. A video that fails none of these checks is not proven real; it is simply not proven fake.
Can you tell an SI video just by watching it?
Not reliably, and the research on this is consistent. The oldest well-known experiment concerns face-swap deepfakes rather than fully generated clips. In a pre-registered study published in iScience in 2021, Nils Köbis, Barbora Doležalová and Ivan Soraperra showed 210 participants 16 short videos, half real and half deepfakes. Average accuracy was 57.6%. Participants labelled 67.4% of the videos as authentic even though they had been told half were fake, and neither a warning about deepfakes nor a cash reward for correct answers improved their scores. They also overestimated their own skill, and those who overestimated it most tended to be the least accurate. The authors call it a "seeing-is-believing" heuristic.
Newer work on text-to-video models tells a similar story. A 2026 research poster by Catherine Davodi, Sarah Barrington, Hany Farid and Emily Cooper reports an online study in which 376 participants judged 480 clips: 24 real videos paired with matching videos from Veo 3.1, Sora 2 and Seedance Pro 1, each shown as a still frame or for 2 to 8 seconds. Accuracy varied a lot between clips and between people. Seedance was the easiest generator to catch and Sora the hardest. Watching longer helped people spot generated clips, but did nothing for real ones, and clips showing humans were easier to flag than clips without them.
The most useful line in that poster is its conclusion: viewers "search for violations rather than verify authenticity". A fake gets caught when something moves wrongly, but a real video has no feature that certifies it as real. As generators improve, the absence of visible errors becomes weaker and weaker evidence.
An eye-tracking study published on arXiv in 2026 by Danqing Shi and colleagues adds a nuance. Forty participants watched real clips and five-second versions regenerated with Veo 3.1 from the same descriptions and keyframes. Mean accuracy was 66.4%, higher than in the online studies, and 36 of 40 people beat chance. In the examples the authors describe, people who caught a fake had looked at the exact spot where it broke: a stick that turned from one into two, or a tennis ball that drifted from green to brown. So looking does help, if you know what to look for and give yourself time.
Step 1: look for a label
Before you study a single frame, check what the platform says. Labels are imperfect, but they are fast.
- YouTube requires creators to disclose realistic content that makes a real person appear to say or do something they did not do, alters footage of a real event or place, or generates a realistic scene that did not happen. Its help page says the label may appear in the expanded description or, for photorealistic content, on the video player itself. YouTube may also add a label on its own for videos made with its own generative tools, for files that carry C2PA metadata, or when its systems detect generated or altered content.
- TikTok announced on 9 May 2024 that it would automatically label SI content made on some other platforms by reading its Content Credentials. TikTok itself warned that the rise in labels would be gradual, because the metadata has to be present for the system to recognise the clip.
A label is strong evidence when it is there. Its absence means very little: disclosure depends on honest creators, and automatic labels depend on metadata that is easily lost on the way.
Step 2: look for a visible watermark
Some generators stamp their output. When OpenAI launched Sora 2 and the Sora app on 30 September 2025, it wrote that all outputs carried a visible watermark at launch, that every Sora video also embedded C2PA metadata, and that it kept internal reverse image and audio search tools able to trace videos back to Sora. The same page now notes that, as of 26 April 2026, the Sora product is no longer available. Clips made with it still exist, though, and can be reposted anywhere.
A visible mark in a corner is worth checking, but treat it as one-way evidence. A logo can be cropped out, blurred or covered by captions when a clip is reposted, so a clean corner proves nothing. Look instead for traces of tampering: an oddly tight crop, a frame that is slightly zoomed compared to other copies of the same clip, or a blurred patch where something used to be.
Step 3: check the file's credentials
If you can get the original file, not a screen recording or a re-shared copy, two kinds of hidden marks may still be inside.
Content Credentials (C2PA) are signed records attached to a file that can name the tool that made it. We explain how they work, and why they are easy to lose, in our guide to C2PA and Content Credentials. Our SI video detector reads them along with other synthetic media flags and recording tags in the file.
SynthID is the invisible watermark Google puts in content from its models, including Veo. According to Google's Gemini help page, you can upload a video to the Gemini app and ask whether it was made with Google AI; files must be 100 MB or less and shorter than 90 seconds, and there is a daily limit of roughly ten video checks. In October 2026, PetaPixel reported that Google had opened its SynthID Detector website to everyone, and that it now also supports content associated with OpenAI, Nvidia and Kakao.
Both checks share the same weakness. A positive result is meaningful. A negative one only tells you that this particular mark is not there: Google's own help page says a video without SynthID may still have been made by another company's model.
Step 4: watch the motion, not the pixels
Single frames from modern generators are often flawless. Errors tend to live in time, which is why participants in the Davodi study mentioned motion more than anything else (66% of them), followed by hands and other body parts, then image quality. Watch the clip at least three times, once at normal speed, once slowed down if your player allows it, and once while ignoring the main subject.
- Object permanence. Do things keep their number, shape and colour? Count legs, fingers, wheels, sticks, buttons. Watch objects that pass behind something and come out again.
- Physics. Does liquid pour and splash the right way? Do footsteps match the ground, does a ball bounce with the right weight, does hair or fabric react to movement and wind?
- Contact. Hands holding cups, feet touching floors, people hugging: the points where two things touch are hard to generate and often slide or merge.
- Text and signs. Shop fronts, number plates and screens may look like writing at first glance and turn into nonsense when paused.
- Sound. If the clip has audio, does the room sound fit the space, and do lips match the words? Generated or cloned voices can sound flat, or too clean for the situation.
- Continuity. A long continuous take where every detail stays consistent from start to end is harder to fake than a few seconds of a single shot. A short clip proves nothing, since most real phone videos are short too.
Our guide on how to spot SI images covers the still-image tells, such as hands, reflections and textures, that also apply to paused frames.
Step 5: trace the source
This step settles more cases than any visual check. Ask who posted the clip first, when, and with what claim. A real event usually leaves several videos from different angles, local news coverage and witnesses. A generated clip often appears from nowhere, on an account with no history, and every copy online traces back to the same file.
Take screenshots of two or three clear frames and run them through a reverse image search; our step-by-step guide on how to reverse image search shows how on a computer and a phone. You may find the original upload, an earlier version with a watermark, or a fact-check that has already done the work.
What about deepfakes of real people?
A face swap or a lip-sync fake is a different problem. Usually the footage is real and only a face or a voice has been changed, so physics and background look perfect. Focus on the face edges, the teeth and the mouth when the person speaks, and on whether the person would plausibly say this at all. We cover the main types and real cases in what is a deepfake. Be honest about your tools too: our video detector analyses whole frames and does not check faces or whether lips match the audio, so it is not a deepfake detector.
We tested three videos with our detector
To see how a frame-based detector behaves on real material, we ran three public domain or CC0 clips from Wikimedia Commons through our SI video detector. All three are re-encoded WebM copies, which is typical of what you find online, and none of them carried Content Credentials, an SI flag or phone recording tags. The verdicts below are the ones a visitor would see.
Veo 3 clip, "Chernobyl decaying room" (8 seconds, with sound, shared on Commons as a Veo 3 example dated August 2025): 99% likely SI-generated. All eight frames leaned SI, from 89% on the first frame to 100% on the last five.
Sora clip, "Borneo wildlife on the Kinabatangan River" (8.5 seconds, hornbills on a branch over a river, credited to OpenAI and dated 15 February 2024 on Commons): 0% likely SI, "Likely real footage". All eight frames leaned real. This is a clear miss.
Real footage, "Nerang River at Numinbah Valley, Queensland" (29 seconds, filmed in January 2016, CC0): 0% likely SI, "Likely real footage". Correct.
The Sora miss is the useful part of this test. Without credentials in the file, the detector has only the pixels to go on, and a photorealistic nature scene with no people, no hands and no text gave it nothing to hold on to. That is exactly the kind of clip that would also fool a human, as the Davodi study found with Sora. One correct catch, one correct pass and one confident wrong answer on three files says nothing about overall accuracy, but it shows why a detector score is one clue among several. If you want to check for SI yourself, upload the original file whenever you can, read the frame scores and the metadata panel, not just the headline number, and combine the result with the steps above.
A five-minute checklist
- Read the post: is there a platform label, a creator disclosure or a credit to an SI tool?
- Check the corners and the edges for a watermark, or signs that one was cropped or blurred.
- If you have the original file, check it for Content Credentials and, for Google tools, SynthID.
- Watch it three times, focusing on motion, contact points, text and sound.
- Find the first upload and look for other angles of the same event.
- Run a detector if you have one, and weigh its verdict with everything else.
- If you still cannot tell, say so, and do not share it as real.
The bottom line
There is no single trick that reveals an SI video, and people who think they have one tend to be overconfident. What works is stacking weak clues until they point in one direction: a label or watermark, credentials in the file, motion that breaks, a source that leads nowhere. Every one of those checks can come back empty on a generated clip, so "I found nothing wrong" should never become "this is real". When a video matters, the source and the context will tell you more than the pixels.



