Guides · Article 50 · EU AI Act
Watermarking vs metadata for Article 50(2) marking
Teams reading Article 50(2) of the EU AI Act often ask a practical question: does 'marked in a machine-readable format' mean watermarking, metadata, or something else? This guide looks at what the regulation text actually asks for and how teams commonly think about the options. It is informational only, not legal advice, and not an endorsement of any technology.
What Article 50(2) actually asks for
The regulation text requires providers of AI systems generating synthetic audio, image, video or text to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. It then describes the technical solutions in terms of qualities, not products: they must be effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities and limitations of various types of content, the costs of implementation and the generally acknowledged state of the art, as may be reflected in relevant technical standards (Article 50(2), Regulation (EU) 2024/1689).
Notably, the text does not mandate one specific technology. It sets an outcome — machine-readable, detectable marking — and qualities the solution should have, leaving the method to track the state of the art.
Watermarking and metadata as common approaches
In practice, two families of technique come up most often, and they are illustrative rather than required:
- Watermarking-style approaches embed a signal in the content itself — patterns in pixels, audio, or token choices — so the marking travels with the content even when metadata is stripped.
- Metadata-style approaches attach provenance information alongside the content — for example in file headers or a signed manifest — which can carry richer detail but may not survive every re-encode or screenshot.
- Many teams consider layering both, since each has different robustness trade-offs across content types.
How teams choose and document
Because the text judges solutions by effectiveness, interoperability, robustness, and reliability so far as technically feasible, the useful exercise is to record why your chosen method fits those qualities for each output type — audio, image, video, and text each have different limitations, which the text expressly acknowledges. Note any technical standards you are tracking, revisit the choice as the state of the art moves, and have qualified counsel review the documentation. Our free scope check can give an informational read on whether Article 50(2) is in play for your system at all.
Common questions
Does the EU AI Act require watermarking specifically?
No single technology is named in the regulation text. Article 50(2) requires machine-readable marking that makes outputs detectable as artificially generated or manipulated, with solutions that are effective, interoperable, robust, and reliable as far as technically feasible. Watermarking and metadata are common ways teams approach that — which fits your system 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.