Brussels: The European Union has spent years writing the world’s most ambitious rulebook for artificial intelligence, and this summer it reached one of the more delicate parts of the job: making machine-made content admit what it is. In mid-June the Commission published a Code of Practice on the marking and labelling of AI-generated content, a document meant to translate the lofty transparency promises of the AI Act into instructions that engineers can actually build against. The obligations it supports become applicable in early August, giving providers a narrow window to get their systems in order. The code addresses two related duties buried in Article 50 of the Act. One requires that AI-generated material carry a machine-readable mark, a hidden signal that detection tools can read even when a human cannot see it. The other requires that certain content, above all the synthetic images, audio and video known as deepfakes, be labelled clearly enough that an ordinary viewer understands they are not looking at a recording of something real. Together they aim at a problem that has crept from novelty to menace: the growing difficulty of telling whether a voice, a face or a quote was produced by a person or a model. What gives the exercise its edge, and its critics, is a single word: voluntary. The code is a guide, not a statute. Companies that follow it gain a presumption that they are meeting their legal obligations, a valuable shield against enforcement, but nothing forces them to sign on, and a firm that ignores it can still argue it has complied with the law by other means. Supporters call this pragmatic, a way to move faster than formal rule-making allows and to keep pace with a technology that mutates by the month. Detractors see a familiar pattern in which the hardest choices are outsourced to the very industry being regulated. The technical obstacles are stubborn. Watermarking schemes can be stripped out by anyone determined enough, cropping and re-encoding erode hidden signals, and there is no single standard that every platform reads the same way. A label is only as good as the weakest link in the chain that carries content from creation to a viewer’s screen, and that chain is long. Skeptics note that the actors most likely to abuse synthetic media, from fraud rings to disinformation operations, are precisely the ones least likely to volunteer a warning label on their work. Still, the effort is not empty. Guidelines published in the spring set out the Commission’s reading of the transparency duties, and together with the new code they form the scaffolding that courts and regulators will lean on when disputes arise. Even a voluntary framework shifts expectations, establishing what responsible behaviour looks like and giving honest firms a template to follow. Whether it meaningfully slows the flood of convincing fakes, or simply lets the diligent label themselves while the reckless carry on, is a test that begins the moment the obligations bite.