Labelling AI images and deepfakes: from 2026


A generated product shot in the catalogue, an avatar in an explainer video, a synthetic voice in an ad: from 2 August 2026 two different obligations attach to these — and they hit two different parties. The provider of the system has to mark the file in machine-readable form. Whoever publishes the content has to disclose it to people. The two together make compliance; either one alone does not.
Two obligations, two addressees
Provider obligation (Art. 50(2))
- Addressee: whoever places the generating system on the market.
- Form: machine-readable marking in the output, detectable as artificially generated or manipulated.
- Standard: effective, interoperable, robust and reliable, as far as technically feasible.
- Applies to audio, image, video and text — including general-purpose AI systems.
Deployer obligation (Art. 50(4))
- Addressee: whoever uses and publishes the content under their own responsibility.
- Form: disclosure to the people who see or hear the content.
- Timing: at the latest at the first exposure, clearly distinguishable (para. 5).
- Applies to image, audio and video constituting a deepfake.
What counts as a deepfake under the regulation
The term is broader than everyday usage suggests. It means not only a faked video of a real person but AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic. A photorealistically generated product shot that looks like a photograph can fall under it — even where nobody was meant to be deceived. The standard is the effect on the viewer, not the intent of the creator.
The most common error in thinking
Many teams rely on “the tool marks it anyway”. But the provider’s watermark satisfies the provider obligation, not yours. It sits in the file and is meant for machines — the person seeing the image on the website learns nothing from it. The reverse holds too: a visible note beneath the image does not replace machine-readable marking.
When the obligation is softened
For evidently artistic, creative, satirical or fictional works and programmes the regulation provides a milder form: there it suffices to disclose the existence of the artificial generation or manipulation in an appropriate manner without hindering the display or enjoyment of the work. End credits can therefore be enough where a watermark across the image would destroy the work. The decisive word is “evidently” — the exception holds where the artistic character speaks for itself, not where it is merely asserted.
For the provider obligation under paragraph 2 the exceptions are technical: systems performing an assistive function for standard editing, and systems that do not substantially alter the input data provided by the user or its semantics, are exempt. In practice that means retouching, cropping, noise reduction and colour correction stay outside. As soon as a tool invents image content — removes a person, replaces a background, adds to a scene — it leaves that territory.
Video and audio: the same provision, more practical questions
The same paragraph applies to video and audio, but implementation is more awkward. An image carries its note in the caption; a 40-second spot requires a decision on when the note appears and for how long, so that it is still “clearly distinguishable” at first exposure. With synthetic voices and real-time avatars there is the added point that disclosure has to happen acoustically or visually before the conversation runs — and that paragraph 1 can apply alongside where people interact with the system rather than merely listening.
The EU symbol and the Code of Practice
The Code of Practice on Transparency of AI-Generated Content was published on 10 June 2026. It is voluntary: providers and deployers sign up to the sections relevant to them via a form submitted to the AI Office. Section 1 concerns providers — marking and detection, expressly subject to technical feasibility. Section 2 concerns deployers and covers the labelling of deepfakes as well as of text published to inform the public on matters of public interest. Annex I provides an optional EU symbol in three variants that deployers can use for labelling.
Final version following a process involving industry, academia, civil society, rightsholders and member states. Signing is voluntary.
From here marking and disclosure apply. Anyone who has signed the Code follows its commitments.
For systems placed on the market before 2 August 2026, the Code allows a transition period until this date.
Voluntary does not mean non-binding
The Code does not replace the obligation under Art. 50 — it describes one way of meeting it. Anyone who does not sign has to meet the obligation another way and, in case of doubt, justify that themselves. Anyone who signs has a documented standard against which they will be measured.
What to do now
- Build an inventory: which tools generate image, video or audio in your organisation — including subscriptions held by individual departments?
- Clarify your role: deployer, or provider yourself through substantial modification? That decides which of the two obligations hits you.
- Walk through the publication channels: website, catalogue, social, newsletter, packaging, trade fair stand. The obligation arises where the content goes out.
- Decide and document the form of disclosure — deciding once per channel beats renegotiating per asset.
- Check whether the provider you use marks its output and whether it has signed the Code of Practice. That enquiry is part of due care towards providers.
Reguly holds this chain together in one place: the AI inventory lists every system with its provider and area of use, the obligations checklist hangs off the individual system with article references, and the provider requests document what you asked the vendor about marking — with status and response rate instead of an email thread in someone’s inbox.
Talk to us
Where does generated content leave your organisation?
We walk through the channels in which AI images, video or audio appear in your organisation and assign the right obligation to each.
Arrange a call



