EU Locks In Its Rulebook for Labelling AI-Generated Content, Eight Weeks Before Transparency Rules Bite

The European Commission has published the final Code of Practice on marking and labelling AI-generated content, closing a drafting process that ran for seven months and involved hundreds of companies, researchers and civil society groups. The timing is tight. The AI Act’s transparency rules begin to apply across the bloc on 2 August 2026, which leaves providers of generative AI systems less than eight weeks to get their houses in order — and the code, released on 10 June, is the closest thing they will get to an instruction manual.

Signing is optional. The code is, in the Commission’s words, a voluntary tool for providers and deployers of generative AI systems to demonstrate compliance with Article 50 of the AI Act. That article is the law’s answer to a problem most internet users now recognise on sight: content that looks real, sounds real, and was made by a machine. Companies that adopt the code get a documented, Commission-blessed route to compliance. Those that go their own way will have to show regulators their alternative works just as well.

What actually changes on 2 August

Three obligations arrive at once. Deepfakes — AI-made images, audio or video that convincingly resemble real people, places or events — must be clearly labelled. So must AI-generated or AI-manipulated text published to inform the public on matters of public interest. And anyone talking to an interactive AI system, a chatbot being the obvious case, must be told they are dealing with a machine.

Beneath the visible labels sits a technical layer for providers. Outputs of generative systems have to be marked in a machine-readable format so that other software can detect them as artificial. The Act asks that these marking solutions be effective, interoperable, robust and reliable as far as technically feasible — a phrase that acknowledges, quietly, that watermarking text is a much harder problem than watermarking an image.

There is one carve-out worth knowing about. AI-assisted text escapes the labelling duty when it has gone through human review and someone carries editorial responsibility for publishing it. A newsroom that uses AI in drafting but edits and answers for what it prints is treated differently from an unsupervised content mill. Where exactly that line falls in practice is likely to be tested quickly.

Seven months, three drafts, two working groups

The AI Office, the Commission body that polices the AI Act, kicked off the drafting at a plenary on 5 November 2025 and split the work along the same seam as Article 50 itself: one working group for providers, who must mark content at the source, and one for deployers, who must disclose deepfakes and public-interest text. Independent chairs steered the sessions. A first draft appeared on 17 December 2025, a second on 3 March 2026, and the Commission layered a public consultation on its accompanying draft guidelines on top, which closed on 3 June.

The final package goes beyond the code itself. The Commission has also released a set of common EU icons for flagging AI-generated content — an attempt to give labelling a consistent visual language rather than a patchwork of in-house badges — along with the signature process for companies that want in. An information session for prospective signatories is scheduled for 22 June.

Swimming against the deregulatory tide

The release lands at an odd moment for the AI Act. Much of the rest of the rulebook is moving in the opposite direction: under the Digital Omnibus on AI, provisionally agreed by the Council and Parliament on 7 May and still awaiting formal adoption, the obligations for high-risk systems in areas such as biometrics, education and employment slip to 2 December 2027, and to August 2028 for AI built into regulated products. The same deal adds a new prohibition, from 2 December 2026, on AI systems that generate non-consensual intimate imagery — the so-called nudification apps — and child sexual abuse material.

The transparency deadline, though, has not moved. Brussels has chosen to give industry breathing room on high-risk paperwork while holding firm on the rule that touches what ordinary people see in their feeds every day. Enforcement carries real weight: the AI Act provides for fines of up to €15 million or 3 percent of worldwide annual turnover for breaches of its transparency obligations.

For companies shipping generative features into the EU, the homework is now well defined. Build machine-readable marking into outputs, wire up visible labels for deepfakes and public-interest text, make chatbots introduce themselves — and decide, before August, whether to sign the code or argue equivalence the hard way. The Commission has published the full code, the icon set and a Q&A on its digital strategy site.