Watermarking
Embedding a hidden signal in generated content (image, video, audio or text) so it can later be detected as AI-generated.
Why it matters. Helps platforms and users identify AI-generated media, but detection usually needs the maker's own tool and is not foolproof.
How it works
An AI watermark is a signal embedded in the content itself, such as slight changes to pixels, audio or word choices, that people usually cannot notice but a detector can find. Unlike metadata, it stays with the content through copying and many common edits.
Google's SynthID is one example, applied to content from Google's models, with detection offered through Google's tools. Watermarks are generally designed to survive common changes such as compression and cropping, but heavy edits can weaken them, and one maker's watermark is not detectable with another maker's detector.
Treat watermark detection as one signal among several. Not finding a watermark does not show content is human-made, only that no watermark the detector knows was found.
Worked example
Illustrative: a news desk runs a viral image through a maker's watermark detector, which flags it as likely made by that maker's model. A second image shows no watermark, which only rules out the models whose watermarks that detector checks for.
Common mistakes
- Concluding an image is real because no watermark was detected.
- Assuming one maker's detector can find watermarks from other makers' models.
- Relying on watermarks alone without also preserving C2PA metadata.
See it on a real card: Safeguards on Veo 3.1.