How to tell if a video is AI generated

Start with the label the platform already applied

• YouTube: creators are required to disclose realistic altered or synthetic content at upload. The disclosure shows in the expanded description, and on the player itself for sensitive subjects such as health, news and elections.
• YouTube, the other label: 'Captured with a camera' in the 'How this content was made' panel. It is not a self-declaration. It appears when the creator shot with C2PA capable hardware or software and the credential checked out.
• TikTok: labels arriving content automatically when it carries Content Credentials, and attaches Content Credentials to video made with its own AI effects. Those credentials stay on the file when it is downloaded.
• Meta: the 'AI info' label on Facebook, Instagram and Threads, applied from the C2PA and IPTC markers that the large generators write into their output.

What the file carries, if you can get the file

A video downloaded from TikTok goes into Content Credentials Verify like an image does, and a manifest, if present, names the tool and lists the edit steps in order.

SynthID is the other readable mark. Google states that its Veo output is watermarked invisibly in the pixels, and the SynthID Detector portal reads it. Coverage stops at Google models: a silent result says nothing about Sora, Runway or Kling.

Then there is capture metadata. Video straight off a phone usually records the make and model of the device in the container, and a file that has been through an upload usually does not. Presence points toward a camera. Absence is what an upload does to everything.

What still shows in the picture

• Length and cutting. Generated clips run short and often hold a single continuous shot, because a cut means a second generation that has to match the first.
• Sustained contact between things. Feet on a floor, a hand on a rail, liquid poured into a glass. Objects slide, drift or pass through each other where contact has to hold for more than a moment.
• Text across frames. A sign or a logo that respells itself as the camera moves.
• People at the back. Faces that reshape, a walk cycle that stutters, a crowd that gains someone between one second and the next.
• Speech. Lips out of step with the audio, or a voice recorded in no particular room.
• Late physics. Hair, cloth, dust and water settling behind the movement that caused them, or not settling at all.

Why video detectors are the weakest tool in the box

Deepfake-Eval-2024 assembled deepfakes that genuinely circulated during 2024: 45 hours of video, 56.5 hours of audio and 1,975 images, gathered from 88 websites in 52 languages. Open-source detectors lost roughly half their area under the curve on the video portion compared with the academic datasets those models are usually scored against.

The free video detectors available to the public are largely those same open models behind a form. That gap between the tested performance and the advertised performance is the reason there is no video check on this site, and the reason a percentage from one of those services should not decide anything for you.

Questions that come up before an argument

What actually settles a video

• Find the earliest posting, not the copy in front of you. Provenance in the file and provenance in the timeline are both worth having.
• Pull a keyframe and reverse image search it. Older footage recut with a new caption is more common than pure generation.
• Look for a second angle. Real events in public get filmed by more than one phone.
• Ask who filmed it. An uploader who cannot produce the original, or who did not shoot it, is the answer in most cases where there is one.