The visible logo is just the surface
If you're using generative video models like Google Veo, OpenAI Sora, Runway, or Kling, you're familiar with the visible corner watermark. Removing it - whether by upgrading to a paid tier (as low as ~$8/month on Kling, up to $200/month on Sora Pro) or using an in-browser tool to inpaint the corner - makes your B-roll look clean and professional.
But there is a widespread misconception that scrubbing out that logo makes the video untraceable to AI detectors. From a digital forensics perspective, the visible logo is nothing more than a superficial overlay. The real tracking happens deep within the pixels.
What is SynthID?
Google DeepMind's SynthID, and similar deep neural watermarking technologies, do not rely on static pixels burned into a corner. Instead of appending metadata (which can be easily stripped), SynthID modifies the latent pixel distributions and frequency domain coefficients across the entire video.
The modifications are kept below the human Just Noticeable Difference (JND) threshold. To your eyes, it looks like standard sensor noise. To a Bayesian detection model, it's an unmistakable statistical signature.
Why cropping or inpainting fails to remove SynthID
When you use an inpainting algorithm or crop the video to remove a corner logo, you alter a tiny localized area—typically less than 1% to 3% of the total canvas.
Because SynthID is distributed redundantly across the entire spatial grid (every pixel of your 1920×1080 frame) and propagated across time (every frame in the clip), over 97% of the watermarked signal survives your edits.
This means that if you submit an inpainted video to an AI detector, it will still reliably flag it as AI-generated. The forensic signature is engineered to withstand:
- Spatial cropping and aspect ratio modifications.
- Heavy H.264 or H.265 lossy compression.
- Frame rate conversions (e.g., 24fps to 30fps).
- Color grading, contrast adjustments, and filters.
C2PA metadata is a different story
While pixel-bound statistical watermarks like SynthID survive editing, container-level metadata like C2PA Content Credentials do not. If you process your video through a browser-based editor that re-encodes the file, the original cryptographic C2PA signature is invalidated.
However, automated platform checks (like YouTube's synthetic content flagging systems) do not just look at C2PA metadata. They read the underlying statistical signals.
What this means in practice
Removing the visual watermark is a cosmetic edit. It cleans up your footage for client work and portfolios - nothing wrong with that. But it won't fool platform AI detectors. The statistical fingerprint survives every edit you throw at it.