Brussels: As the deadline approaches for Europe’s most contested technology law, the European Commission has handed companies a practical guide to one of its thorniest obligations: how to mark and label content generated by artificial intelligence. The Code of Practice, published this month, is meant to help providers and professional users comply with the transparency rules of the AI Act before they become binding in early August, when machine-made text, images, audio and video must carry signals that reveal their artificial origin.
The rules spring from Article 50 of the AI Act, the bloc’s landmark attempt to govern artificial intelligence according to the risk a system poses. The transparency provisions require providers to embed machine-readable markers in AI-generated or manipulated content, and oblige those who deploy generative tools for professional purposes to label deepfakes and AI-written material on matters of public interest. The intent is to let citizens know when they are looking at something a machine produced, a safeguard against the erosion of trust that convincing synthetic media threatens.
The Code of Practice does not change the law; it translates it into workable terms. Drafted with input from industry, civil society and technical experts, it sets out how watermarking, metadata and disclosure can satisfy the obligations without strangling legitimate uses of the technology. That distinction matters, because the same tools that fabricate disinformation also power benign applications in film, design, accessibility and journalism. The challenge for regulators has been to flag the deceptive without burdening the creative, and the code represents the bloc’s attempt to draw that line in practice rather than principle.
The stakes are high and the timing tight. From 2 August, the transparency rules become applicable and the Commission’s own enforcement powers enter into force, backed by the European AI Office and national authorities. Companies that have spent months arguing the rules are vague or premature now have a reference document against which to measure their systems, though many will complain it arrives late, leaving only weeks to retrofit products built without labelling in mind.
The wider context is a fierce debate over whether Europe is regulating artificial intelligence wisely or hobbling its own industry. Critics, including some governments and a chorus of technology firms, warn that the AI Act’s compliance burden could push talent and investment toward jurisdictions with a lighter touch. Defenders counter that trust is the precondition for adoption, and that a market flooded with unlabelled synthetic content would ultimately serve no one, least of all the companies building the tools.
Marking content is also only as effective as its weakest link. Watermarks can be stripped, metadata can be discarded, and bad actors operating outside the bloc face no obligation to cooperate. The code acknowledges these limits, framing labelling as one layer of defence rather than a cure. Its success will depend on whether platforms preserve the signals as content travels across the internet, and whether detection tools can keep pace with the generators that grow more sophisticated by the month.
For now, the publication marks a transition from drafting to doing. The AI Act has long been a statement of intent; with August looming, it is becoming a set of demands that companies must meet, and the labelling of synthetic media is the first test of whether Europe’s rulebook can survive contact with the technology it seeks to tame.




