Digital watermarking embeds a marker into an image, video, audio file, or text that can later be detected to show where the content came from or that an AI system generated it. The marker can be visible, like a logo, or hidden in the content itself.

‍

‍Why it matters

For trust and safety teams, watermarks help with provenance, meaning where content came from. They can help flag AI-generated media, trace leaked material, and support takedowns of copies. The limits matter. Watermarks can be weakened by cropping, compression, or re-encoding, and plenty of AI-generated content carries no watermark at all. Finding one is useful evidence. Not finding one proves nothing.

‍

‍How moderation teams use it

A detected watermark is treated as one signal in a review, not a decision. A moderator still applies the platform's own policy on synthetic media, which might require labeling instead of removal. Teams also keep in mind that bad actors have every reason to remove marks.

‍

‍Watermarks versus provenance metadata

These get lumped together but work differently. A watermark is built into the content signal, such as the pixels, which makes it harder to strip. Provenance metadata is a signed record attached to the file, like the Content Credentials approach from the C2PA standards group, and it can be lost if the file is re-saved or stripped. Google DeepMind's SynthID is an example of a watermarking system for AI-generated content. Many teams look at both kinds of signals together.