TrueSight logo TrueSight.
Scan Upload The Wall FAQ Support

Sora & Veo AI Video Detector — How to Tell If a Video Was Made by AI

Text-to-video generators like OpenAI Sora, Google Veo, Kling and Runway now produce clips that look like phone footage. Here is what they still get wrong, and how to check a video you're unsure about.

Last updated: September 2026 • TrueSight (tsdetect.com)

A few years ago, AI video was easy to spot: melting faces, six-fingered hands, a dream-like wobble. Modern generators have fixed most of that. A short Sora or Veo clip of a street interview, a news event or an animal doing something funny can pass a casual glance, and these clips are now all over TikTok, Reels and Facebook.

The good news is that generators still leave consistent, repeatable artifacts. They're just harder to see in one frame and easier to see across many.

Why the watermark isn't enough

Several AI video tools mark their output. Sora clips can carry a visible watermark and embedded provenance metadata (Content Credentials, also called C2PA), and Google says Veo output carries an invisible SynthID watermark. Those are genuinely useful when they survive. The problem is that they usually don't:

  • Cropping removes a corner watermark in one step, and watermark removers are widely available.
  • Re-uploading to a social platform re-encodes the video, which commonly strips the metadata.
  • Screen recording produces a brand-new file with none of the original provenance.

So the absence of a watermark tells you almost nothing. Treat a watermark as a confirmation when you find one, never as proof of authenticity when you don't.

The tells AI video generators still leave

1. Text that won't hold still

Signs, license plates, shop names, jersey numbers and captions inside the scene are the most reliable giveaway. Generators draw letter shapes rather than spell words, so text is often garbled, or readable in one frame and different in the next.

2. Objects with no permanence

Watch the background, not the subject. A cup that changes shape, a person who walks behind a pole and doesn't come out, a car that merges into another car. Generators model what looks plausible frame to frame, not a consistent world.

3. Physics that's almost right

Liquids that pour but never fill, fabric that moves without wind, footsteps that don't quite meet the ground, shadows that don't match the light source. These are subtle at full speed; step through the video slowly.

4. Hands, teeth and crowds

Hands are much better than they were, but still break during fast gestures or when holding objects. Teeth can blur into a single strip. Background faces in crowds are frequently smeared or identical.

5. Too clean, too short, too perfect

Many generated clips are short, perfectly lit, perfectly framed and have a single continuous shot with no cuts. Real phone footage is shaky, badly exposed and interrupted. A "lucky" camera angle on a dramatic event is a reason to look closer.

6. Audio that's synced but hollow

Newer models generate sound too. Listen for voices with no room echo, crowd noise that doesn't match the crowd, and speech that lines up with the lips but sounds studio-clean in a noisy street.

How to check a suspected Sora or Veo video

  1. Check the source first. Who posted it originally? Is there any other angle of the same event from a different account? Real events get filmed by more than one person.
  2. Scan it. On desktop, the free TrueSight Chrome extension captures frames from the video while it plays. On any device you can paste the link for a quick check.
  3. If the result is Unclear, go to the file. Streams and screen captures are compressed, and compression erases exactly the fine detail that gives generators away. Uploading the video file lets us analyze up to 24 frames at full quality.

Frequently Asked Questions

Can you tell if a video was made with Sora?

Often, yes, but not from the watermark alone. Sora videos can carry a visible watermark and embedded provenance metadata, but both are easily lost when a clip is cropped, re-encoded or screen-recorded. The more reliable approach is to look for generator artifacts across many frames: garbled text, objects that appear or vanish, unnatural physics and inconsistent hands or faces.

Does removing the watermark make an AI video undetectable?

No. A watermark is a label, not the evidence. Removing it removes the easiest clue, but the visual artifacts produced by the generator are still in the frames. Frame-based analysis looks at those artifacts rather than the label.

What is the most accurate way to check a suspected AI video?

Analyze the original video file rather than a compressed stream or a screenshot. Every re-upload and screen capture destroys fine detail. Uploading the file lets TrueSight analyze up to 24 frames at full quality.

Is any AI video detector 100% accurate?

No. Generators improve constantly and no detector catches everything. TrueSight gives a verdict with a confidence score and says Unclear when the evidence is weak, rather than guessing. Combine it with checking the source of the video.

Honest limits: TrueSight estimates the likelihood that a video is AI-generated from visual signals. It doesn't identify which generator made a clip, and it can't read invisible watermarks such as SynthID. For anything consequential, pair the result with source verification.

Check a video now

Free link scan in seconds, or a full forensic scan of the file from $2.99.

📤 Upload a video for a forensic scan

Or scan a link free — no account needed.

Related guides