From 2 August 2026, providers of AI systems that generate synthetic audio, image, video or text must ensure the output is marked in a machine-readable format and detectable as artificially generated or manipulated. This follows from Article 50(2) of the AI Act and also applies to general-purpose AI systems (GPAI). The duty sits with the provider, not the party deploying the system. Generative systems that were already on the market before 2 August 2026 have a backstop: they have until 2 December 2026 to meet the marking requirement.
This page accompanies our practical overview of Article 50 and our explainer on the Code of Practice on transparency of AI content. Below you will find what the marking duty entails, who carries it, which techniques qualify and which exceptions apply.
What does machine-readable marking entail?
Article 50(2) requires providers to ensure that the output of their generative AI system is marked in a machine-readable format and detectable as artificially generated or manipulated. This is not a visible label for the viewer, but a signal that machines can read. The solutions must be effective, interoperable, robust and reliable as far as this is technically feasible, taking into account the specific features and limitations of the different content types, the costs and the state of the art.
That nuance matters. The law does not demand perfection, but it does demand a serious effort that fits what is technically reasonable for the content type concerned. Audio, image, video and text each have their own possibilities and limitations.
Who must mark?
The duty sits with the provider: the party that develops the generative AI system and places it on the market, including providers of general-purpose AI systems. Not the organisation that subsequently deploys the system, and not the end user.
| Question | Answer |
|---|---|
| Who carries the duty? | The provider of the generative AI system |
| For which output? | Synthetic audio, image, video and text |
| From when? | 2 August 2026 |
| And existing systems? | Backstop until 2 December 2026 for systems already on the market |
The misconception that marking is a matter for the party publishing the content is exactly where organisations go wrong. Marking at the source is a design choice for the provider. Anyone integrating an external model into their own service is well advised to set out contractually that the provider supplies this marking.
Which techniques qualify?
The law does not mandate any single technique, but lists a range of options: watermarks, metadata identification, cryptographic methods to prove provenance and authenticity, logging and fingerprints. In practice a layered approach works best, because individual techniques can be circumvented or are limited in scope.
In practice, marking falls short when it:
- consists only of metadata that is lost on saving or sharing;
- consists of a watermark that disappears after a simple edit;
- is not interoperable and is therefore not recognised by common detection tools;
- is not robust against normal processing of the file;
- does not align with a common standard.
The European Commission is developing a Code of Practice on marking and labelling of AI content, with a standardised label and a distinction between fully AI-generated and AI-assisted material. Following that line makes it easier to demonstrate that the marking complies. The Code is not yet final, but it sets a clear direction.
Which exceptions apply?
The duty is not absolute. It does not apply to AI systems performing an assistive function for standard editing that do not substantially alter the input or its semantics, such as a spelling or grammar corrector. There is also an exception for systems authorised by law to detect, prevent or investigate criminal offences.
These exceptions are powerful, but they require discipline. Anyone relying on the exception for assistive editing must be able to explain why the system does not substantially alter the content.
Machine-readable marking versus visible deepfake disclosure
Two duties are often confused. The machine-readable marking under Article 50(2) sits with the provider and targets a signal that machines can read. The visible deepfake disclosure under Article 50(4) sits with the deployer and targets the human who sees the content. Anyone integrating an external model into their own service may face both, and is well advised to set out contractually who provides what.
What should you do now?
Determine whether you are a provider
Determine whether your organisation is a provider of a generative AI system that produces synthetic audio, image, video or text, including an own or further-developed GPAI model.
Choose a layered marking approach
Choose a layered marking approach that is effective, interoperable and robust, and that aligns with a common standard. Combine, for example, a watermark with provenance data instead of relying on a single technique.
Set contracts and keep evidence
When integrating an external model, set out contractually that the provider supplies the machine-readable marking, and keep evidence of the chosen solution and the state of the art you rely on.
Organisations that arrange this before the summer will not have to improvise on 2 August 2026. Treat 2 December 2026 as a backstop for existing systems, not as the default deadline.
Frequently asked questions about machine-readable marking
Practical questions about the marking duty for providers under Article 50(2) of the AI Act.