Guides · Article 50 · EU AI Act
How to watermark AI-generated images
If your product generates images with AI, Article 50(2) of the EU AI Act is the paragraph that addresses marking those outputs. This guide walks through how teams commonly approach image watermarking and what the regulation text actually asks for. It is informational only, not legal advice, and not an endorsement of any watermarking product.
One thing to fix up front: the marking duty in Article 50(2) falls on the provider of the generating AI system — roughly, the party that develops or places the generator on the market — not on every person who happens to share an image.
What Article 50(2) asks for
The regulation text provides that providers of AI systems generating synthetic audio, image, video or text must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, using technical solutions that are effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities and limitations of the content type, the costs of implementation, and the generally acknowledged state of the art (Article 50(2), Regulation (EU) 2024/1689).
Read closely, that sets an outcome — machine-readable, detectable marking — and a set of qualities the solution should have. It does not mandate one technology. Image watermarking is one common way teams pursue that outcome, not a method the Act names.
Approaches teams weigh for images
Because images are re-encoded, cropped, and screenshotted, robustness across those transforms is usually the deciding factor. These are illustrative approaches, not requirements:
- Pixel-domain watermarks embed a signal in the image itself, so the marking can survive some edits even when file metadata is stripped.
- Metadata or manifest approaches attach provenance data alongside the image; they can carry richer detail but may not survive every re-encode or screenshot.
- Layering both is a pattern some teams consider, since each carries different robustness trade-offs for images.
What providers commonly document
- That image outputs are marked in a machine-readable format and detectable as artificially generated.
- Why the chosen method is effective, interoperable, robust, and reliable so far as technically feasible for images specifically.
- Any technical standards you are tracking as the state of the art moves.
- If relying on the assistive-editing or no-substantial-alteration carve-out, the rationale, for counsel's review.
Common questions
Does Article 50(2) require a specific image watermark?
No single technology is named in the regulation text. Article 50(2) asks for machine-readable marking that makes outputs detectable as artificially generated, with solutions that are effective, interoperable, robust, and reliable as far as technically feasible. Watermarking is one common way teams approach that; whether it fits your generator is fact-specific — confirm with qualified counsel.
See what may apply to your business
Answer seven quick questions for an automated, informational indication of which Article 50 obligations appear likely to apply — free, and not legal advice.