Watermarking and disclosure requirements for AI-generated content
A distinct but related set of rules is emerging across multiple jurisdictions specifically around synthetic media: content that is generated or substantially altered by AI, most prominently images, audio, and video. The policy goal is straightforward to state: as generative tools make convincing fake content cheap and easy to produce, regulators in multiple places are moving toward requiring that such content be identifiable as AI-generated, either through visible labeling, embedded metadata, or technical watermarking. Provisions along these lines show up in discussions of the EU AI Act's transparency obligations for certain generative systems, and multiple US states have passed or proposed laws specifically targeting deepfakes, particularly around election-related content and non-consensual explicit imagery, which have been some of the most politically urgent use cases driving this category of law.
The mechanisms proposed tend to fall into two categories: disclosure at the point of generation, where a tool is required to label or watermark what it produces, and disclosure at the point of distribution, where a platform is required to label or restrict content that appears to be synthetic regardless of where it came from. These aren't mutually exclusive, and a builder shipping a generative media tool may face obligations from both directions: obligations on the tool itself as a provider, and separate obligations on any platform that later distributes the output.