Risks & safety
AI Watermarking
Also known as: content provenance
Hidden signals or labels embedded in AI-generated content so it can later be identified as AI-made.
AI watermarking is a way to mark AI-generated content, an image, a piece of text, a video, so it can be recognised as AI-made later, even if it looks completely natural. The goal is transparency: helping people and platforms tell human-made content from AI, which matters for fighting deepfakes and misinformation. It is a promising tool, but not foolproof: determined people can often remove or defeat watermarks.
Watermarking embeds a detectable signal into AI-generated content to mark its origin. It comes in two broad forms: - Invisible/statistical: a subtle pattern the model bakes into its output (pixels, or word choices) that special detectors can read, but humans cannot see. - Provenance metadata: attached, often cryptographically signed, information recording that content is AI-generated (content-credential standards).
Why it matters
The limits
Watermarks can be weakened or stripped (editing, paraphrasing, cropping), and different providers use different schemes, so it is one layer of defence, not a complete solution.
AI watermarking embeds a recoverable signal marking content as machine-generated, spanning statistical output watermarks (e.g. biasing the token sampling distribution toward a secret green-list, detectable with a key; Kirchenbauer et al. 2023), signal-domain watermarks for images/audio, and provenance metadata standards (C2PA content credentials with cryptographic signing).
The core tension is robustness vs quality/impact: stronger watermarks resist removal but may perturb output; text watermarks in particular are vulnerable to paraphrasing and mixing, and cross-provider standardisation is incomplete. Attacks include editing, cropping, re-encoding and adversarial removal, so watermarking complements rather than replaces detection and out-of-band verification for deepfake defence. Governance is a major driver: transparency/labelling obligations for synthetic media (e.g. EU AI Act) push adoption, making provenance an AI-ethics and platform-integrity priority even as technical guarantees remain partial.