detect·deepfakesby Resemble AI
Glossary

Watermarking

Also: AI watermark · invisible watermark · digital watermarking

The practice of embedding a detectable signal — usually statistical, sometimes perceptible — into AI-generated media at generation time, so the content can be identified as synthetic by a downstream detector that knows the watermarking scheme.

Watermarking is the AI-native counterpart to C2PA. Where C2PA adds signed metadata, watermarking embeds a detectable statistical signal in the pixels or audio samples themselves. The goal is the same: identify synthetic content downstream. The tradeoffs are different.

How watermarking works

Two broad approaches:

Perceptible watermarks — a visible logo, a text caption, or an audio bleep that announces the content is AI. Cheap to implement, trivial to remove.

Statistical watermarks — a pattern embedded in the pixel or frequency domain that a dedicated detector can recognize but humans cannot see. Robust to mild edits, fragile to aggressive re-encoding.

Recent examples: Google's SynthID for images/audio, OpenAI's text watermarking experiments, Meta's video watermarks.

Why watermarking matters

Watermarking shifts the detection problem from "does this look AI?" to "does this carry our watermark?" — which is a simpler, more reliable question when the watermarking scheme is known.

China's 2023 regulation requires watermarking on all AI-generated content in-country. The EU AI Act's 2026 labeling requirements push in a similar direction.

Limitations

  • Adversarial removal. Attackers can re-generate content through other models, apply heavy transformations, or use watermark-removal tools.
  • Ecosystem fragmentation. No universal watermarking standard — each model family uses its own scheme, and detectors need the specific scheme to verify.
  • Doesn't help with untrustworthy models. A malicious generator simply won't watermark its outputs.

Relationship to deepfake detection

Watermarking, C2PA provenance, and deepfake detection are three complementary layers:

  • Watermarking catches content from cooperating models.
  • C2PA catches content from cooperating tools.
  • Deepfake detection catches content from everything else.

Most real deployments use all three.

See also