Implementation guide
OWASP AISVS v1.01
Model Behavior, Output Control & Safety Assurance · Source Attribution & Citation Integrity
Watermark generated media
Verify that generated media is watermarked to prove it was AI-generated.
OWASP AISVS v1.01, requirement 7.4.4
In plain English
Mark generated images, audio and video so they can be identified as AI-generated. This helps people and platforms tell synthetic media from real, and several laws now expect it, including the EU AI Act's transparency rules. This is a Level 3 control.
What good looks like
- Generated media carries a watermark embedded in the content, signed provenance data, or both.
- The watermark survives common edits such as resizing, cropping and compression.
- There is a working way to check a file for the mark.
- The marking is applied on the server and cannot be switched off by a user.
How teams usually do it
- Content Credentials using the C2PA standard, which attach signed provenance metadata to the file.
- An embedded watermark from your model provider (for example Google's SynthID) or an open implementation, for the cases where metadata is stripped.
- Both together. Metadata gives detail, and the watermark survives re-uploading.
Suggested evidence
- Generation pipeline configuration.
- A check of sample output for both the credential and the watermark, before and after resizing and recompression.
Common mistakes
- Using metadata alone, which most sharing platforms and any screenshot remove.
- Marking images and forgetting audio and video.
- Treating a watermark as proof. Detection has false positives and negatives, and determined attackers can remove marks.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.