DeepFake Check

Sora, Veo, Kling

Is this video AI? Sora, Veo, Kling and other generators

A clip from Sora, Veo, Kling, or any other generator is still just a video file. This page does not have a separate detector for each name, and a result does not say which product made the clip.

What the video check does

The browser picks up to 10 frames spaced across the whole clip and keeps each frame’s timestamp. Those frames go to a face-forensics service, and Google Gemini reviews the frames for repeated visual conflicts. If the face service fails, the Gemini review still runs. The check does not look at every frame, so a change that falls between samples can be missed.

What it does not do

  • It does not name Sora, Veo, Kling, or any other generator.
  • It does not decode SynthID. Google’s own Gemini tools are what read that watermark.
  • It does not read a C2PA credential from the video file. The credential reader on this site is for images.
  • A low-risk result does not prove the clip, the speaker, or the event is real.

Checks worth doing yourself

  • Watermark. OpenAI’s Sora launch post (30 September 2025, still on their site on 29 October 2026) says every Sora video included a visible watermark and embedded C2PA metadata. The same page says the Sora product is no longer available as of 26 April 2026. A screen recording or a re-export can drop both marks. A missing mark does not mean the clip is real. Source: openai.com/index/launching-sora-responsibly.
  • SynthID on Veo. Google DeepMind’s Veo page says videos made with Veo are marked with SynthID. This check does not read SynthID. Source: deepmind.google/models/veo.
  • C2PA on Veo. Google Cloud’s Content Credentials documentation, read on 29 October 2026, lists Veo 3 and Veo 3.1 model ids among models whose images or videos are signed by Google LLC. Use a C2PA verifier on the original file. This video upload does not do that read. A credential says who signed a claim. It does not prove the depicted event happened. Source: Google Cloud content credentials.
  • Physics, text, hands, and audio. Watch whether objects pass through each other, whether on-screen text stays stable, whether hands keep a plausible shape, and whether speech stays locked to the mouth. One odd frame is not proof. Repeated breaks are a reason to slow down. These are viewing habits, not a measured accuracy rate.
  • Kling. Use the same viewing habits. This page does not state a Kling watermark or C2PA policy, because no Kling vendor page was checked for this update.

Upload the file on the video detector. The method and the limits of the image pixel model are on the accuracy page. That page is not a video accuracy test.

FAQ

Is this video AI if it came from Sora, Veo, or Kling?

Upload the video file. The check samples up to 10 frames and does not say which generator made the clip. A low-risk result does not prove the clip is real. OpenAI has said Sora outputs carried a visible watermark and C2PA metadata, and that the Sora product ended on 26 April 2026. Google says Veo videos are marked with SynthID, which this check does not read.

Does the video check read C2PA or SynthID?

No. SynthID is not decoded here. The C2PA reader on this site is for images, not for the video upload. A missing watermark or credential is not evidence that the clip is real.

What should I look at myself in a generated clip?

Look for a visible watermark, unstable on-screen text, hands that change shape, objects that pass through each other, and speech that slips off the mouth. One odd frame is not proof. These are viewing habits, not a measured accuracy rate. For a Kling watermark or C2PA policy, this page does not quote one.