Watermarking

Watermarking marks SI-generated images, audio, video or text so that software can later tell they were machine-made. It’s one of the main technical responses to deepfakes.

Approaches

• Invisible watermarks: subtle patterns in pixels or audio that survive common edits. Google’s SynthID is a widely deployed example.
• Text watermarks: biasing word choices in a statistically detectable way. These are harder to make robust.
• Content credentials (C2PA): cryptographically signed metadata recording how a file was created and edited, supported by camera makers, Adobe, Microsoft, OpenAI, Google and others.

Limits

Watermarks can be weakened by cropping, compression, screenshots or paraphrasing. Content made with open-source tools may carry no watermark at all. Metadata is easily stripped. Experts see watermarking as one layer among several, not a complete fix.

In law

The EU AI Act requires providers of generative systems to mark synthetic content in a machine-readable way, as part of transparency duties phasing in from 2026. Several U.S. states have disclosure rules for election deepfakes.