AI content watermarking

AI content watermarking embeds machine-detectable signals in AI-generated text, images, or audio to enable provenance verification. Required for some content categories under the EU AI Act Article 50.

What is AI watermarking?

AI content watermarking embeds machine-detectable signals into AI-generated content to enable downstream verification of provenance. For text, watermarking adjusts token-selection probabilities in a way that statistical tests can later detect. For images and audio, watermarking modifies the output in ways imperceptible to humans but recoverable by detector models. Google's SynthID is the most-publicized production text watermarking; C2PA (Coalition for Content Provenance and Authenticity) provides an open standard for signed content metadata.

Regulatory drivers

EU AI Act Article 50 (transparency obligations) requires AI-generated text intended to inform the public on matters of public interest to be labeled as such; AI-generated or manipulated audio, image, and video (deepfakes) must be marked machine-readable. China's Provisions on the Administration of Deep Synthesis Internet Information Services (March 2023) require similar labeling. The US has no federal mandate but several states and the AI Bill of Rights principle of "notice and explanation" pull in the same direction.

Limits

Watermarking is an arms race. Sufficient post-processing — translation, paraphrasing, recompression — degrades or removes watermarks. Open-weight models can be modified to remove watermarking code entirely. For procurement, ask: which content categories are watermarked, how robust is the scheme to common transformations, can the watermark be removed in your deployment (custom fine-tuning, post-processing pipelines), and what is the vendor's policy if a user circumvents watermarking.