On 10 June 2026, the European Commission published the final Code of Practice on marking and labelling of AI-generated content. The Code is voluntary, and sets out practical steps to help the providers and deployers of generative AI systems meet the transparency obligations of the AI Act, which apply from 2 August 2026. It is the first attempt by a major regulator to specify, in operational terms, how synthetic content should be made recognisable as such.
The obligations the Code supports are narrow but consequential. From August, deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest must be clearly labelled. People must also be told when they are interacting with a chatbot or other interactive AI system. The stated purpose is to help the public recognise when content has been generated or altered by a machine, and so reduce the scope for deception.
Two problems, two halves
The Code divides the task in two. One half, drawn from Article 50(2) of the Act, addresses marking and detection: the obligation on those who build generative systems to embed signals in their output, such as watermarks or metadata, that allow a machine or a person to identify it later. The other half, from Article 50(4), addresses labelling: the obligation on those who deploy such systems to show users, plainly, that what they are looking at was produced by AI.
The distinction is not academic. Marking is a property of the file, applied at the point of creation and meant to travel with it. Labelling is a property of the presentation, applied at the point a person encounters the content. The first is invisible and durable in principle. The second is visible and depends on whoever publishes the material choosing to display it. The Commission has also issued a set of EU icons so that labels look consistent across platforms.
The limits the Code carries
A voluntary code is an instrument with a known ceiling. Signing it offers providers a presumption that they are meeting their obligations, but it cannot compel those who decline to sign. Technical marking, meanwhile, is only as strong as its resistance to removal, and watermarks embedded in text or images can be stripped, cropped or paraphrased away. Detection, the other side of marking, remains an unsolved problem at scale.
There is also the matter of what falls inside the rules. The duties bite hardest on deepfakes and on AI-generated text about public-interest matters, the cases where deception does the most damage. A great deal of ordinary synthetic content, the routine drafting and image-making now woven through daily work, sits outside the sharpest obligations. The Code is a floor for the highest-risk uses, not a comprehensive account of where machine-made material now appears.
If the value of a label rests on people being able to trust it, what happens to that trust when the most consequential fakes are precisely the ones most likely to evade marking?
Opinion: Provenance Becomes Infrastructure
The significance of the Code is not in its detail but in what it concedes. It accepts that the line between human and machine-made content can no longer be assumed, and must instead be actively maintained. Provenance, once a quiet by-product of how things were made, is becoming a piece of public infrastructure that has to be built, funded and policed.
That is a reasonable response to a real problem, and the focus on deepfakes and public-interest text is well chosen. The harm from a fabricated statement attributed to a public figure is not symmetrical with the harm from an AI-drafted email, and the rules sensibly concentrate where the stakes are highest. A common labelling standard, with shared icons, is the kind of unglamorous coordination that regulators are well placed to provide.
The harder question is what a label can carry. A mark tells a reader that content is synthetic; it does not tell them whether it is true, fair or manipulative. As synthetic media becomes ordinary, the risk is not only that fakes go unmarked, but that marking becomes background noise, a notice people stop reading, in the way cookie banners taught a generation to click past consent. Transparency is a precondition for judgement, not a substitute for it.
So the open question is whether labelling can bear the weight now being placed on it. Europe has built the first scaffolding for telling human from machine. Whether that scaffolding holds will depend less on the watermark than on whether the institutions and habits that turn information into trust are rebuilt alongside it.
Declaration of Generative AI and AI-assisted technologies in the writing process:
The author made use of Generative AI or AI-assisted technologies in the preparation of this post.
Sources
European Commission, "Commission publishes Code of Practice on marking and labelling AI-generated content," 10 June 2026
European Commission, "Code of Practice on marking and labelling of AI-generated content" (policy page)
European Commission, "EU Icons for labelling AI-generated content"
The contents of this article are for informational purposes only and do not constitute professional, legal, or financial advice.
