AI Watermarking

What Is AI Watermarking?

AI watermarking is the practice of embedding imperceptible signals — watermarks — into AI-generated content at the time of creation. These signals are designed to survive common transformations (editing, compression, reformatting) so that the AI origin of the content can be detected later.

AI watermarking is different from C2PA provenance metadata — watermarks are embedded in the content itself (pixel values, token distributions), not attached as separate metadata that can be stripped.

Types of AI Watermarks

Statistical Watermarking (Text)

During text generation, the model's token sampling distribution is subtly biased using a cryptographic key. Over many tokens, this creates a statistically detectable pattern. Google's SynthID uses this approach for Gemini-generated text.

Perceptual Watermarking (Images / Audio)

For images and audio, watermarks are embedded in frequency-domain representations — imperceptible to human senses but detectable algorithmically. These watermarks are trained to survive JPEG compression, color grading, cropping, and format conversion.

Regulatory Requirements

AI watermarking is no longer optional for major AI providers:

NIST AI 100-4 (US): voluntary guidance on watermarking approaches, referenced in the 2023 Executive Order on AI Safety as mandatory for federal AI systems
EU AI Act Article 50: requires generative AI providers to implement technical measures to mark AI-generated output — watermarking is the primary implementation path
Voluntary commitments: major AI labs (OpenAI, Google, Anthropic, Meta) have committed to watermarking as part of the White House AI Safety agreements

See our full standards overview for regulatory detail.

Limitations

No watermarking scheme is fully robust. Statistical text watermarks degrade after heavy paraphrasing. Perceptual image watermarks can be removed via adversarial perturbation. Watermarking is one layer of a defence-in-depth approach — not a complete solution to synthetic media detection.