EU Publishes First Draft Code of Practice for Labelling AI-Generated Content

EU Publishes First Draft Code of Practice for Labelling AI-Generated Content

Europe is about to make AI-generated content announce itself. As the clock ticks toward the August 2, 2026 application date for the EU AI Act's transparency rules, the European Commission has been steadily building out the machinery that will require deepfakes, synthetic images, AI-written articles and chatbots to carry visible disclosures — and at the center of the effort is a proposed "common icon," a single EU-wide visual cue meant to tell a citizen, at a glance, that what they are looking at was made or manipulated by a machine.

The vehicle is a voluntary Code of Practice on the marking and labelling of AI-generated content, drafted by independent experts under the Commission's AI Office. The first draft was opened to participants and observers late last year, with a stakeholder feedback window that ran to January 23, and a second draft was prepared by mid-March. The Commission has said it expects to finalize the Code in June 2026, just weeks before the underlying legal obligations bite. Running alongside the Code, the Commission published draft guidelines on the implementation of Article 50 on May 8, 2026, whose consultation period closed June 3.

Together, the two documents flesh out what had been one of the AI Act's most abstract promises: that people interacting with AI, or consuming its output, should know it.

What the rules require

The obligations flow from Article 50 of Regulation (EU) 2024/1689, which sets four distinct transparency duties along the AI value chain. Providers of interactive systems must design them so users are told they are dealing with a machine and not a person. Providers of generative systems must mark synthetic audio, image, video or text in a machine-readable format — think watermarks and metadata — so outputs are detectable downstream. Deployers of emotion-recognition and biometric-categorization systems must inform the people exposed to them. And deployers of deepfakes, plus those publishing AI-generated text to inform the public on matters of public interest, must disclose the artificial origin of that content.

The Code translates those legal categories into design specifics. It sets out rules on the placement, size and readability of labels, accessibility standards, and modality-specific requirements for text, images, audio, video and live or real-time content. Its headline proposal is the common icon. To avoid what one analysis called a fragmented landscape of competing warning labels across 27 member states, the draft proposes a uniform EU mark. Until that interactive icon is finalized, signatories may rely on an interim two-letter acronym — "AI" in English, "KI" in Germany, "IA" in France — placed visibly and consistently at the point of first exposure, while committing to support the rollout of the common version.

The Commission's interpretation of who is covered is expansive. In its draft guidelines, it makes clear that a deepfake must be labelled even when no one intended to deceive, and even when the depicted person does not actually exist. "Content that looks or sounds like a real person must be labeled — even if no deception was intended and even if no real individual is depicted," Greenberg Traurig attorneys Philip Radlanski and Carsten Kociok wrote in a June 8 analysis of the guidelines, summarizing the Commission's reading of Article 50(4). Plainly unrealistic content — fantasy scenes, flying humans, dragons — would generally fall outside the definition.

On chatbots, the Commission is equally firm that disclosure must be upfront and obvious. "Users need to know they are interacting with AI at the moment of contact — not through a reference buried in fine print," the Greenberg Traurig lawyers noted, adding that the duty extends to agentic AI that acts autonomously on a user's behalf. Burying the disclosure in terms and conditions, or relying on metadata alone, will not satisfy the rule.

Why the EU is doing this its way

The approach marks a sharp divergence from the United States, where there is no comprehensive federal AI transparency mandate and content-labelling has been left largely to platform policies and voluntary industry commitments. Brussels is instead hard-coding disclosure into binding law, with the Code of Practice serving as a structured presumption of compliance. Adherence to a Code deemed adequate by the AI Office offers signatories a route to demonstrating they meet Articles 50(2), (4) and (5); non-signatories can still comply by other means but, per legal analysts, may face heavier evidentiary burdens and more frequent information requests from market surveillance authorities.

The stakes are not trivial. Infringement of the transparency obligations can draw penalties of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For the largest AI providers, that turnover-based figure dwarfs the fixed cap.

The reach into newsrooms and corporate communications has drawn particular attention. The exemption for AI-generated text that undergoes "human review and editorial responsibility" is being read narrowly: a spell-check or a cursory sign-off will not do. Organizations must show genuine, substantive review by a qualified person and clearly attributable editorial accountability, with a named, contactable party empowered to approve, amend or reject the content. Marketing teams using synthetic visuals and communications departments producing AI-drafted investor materials are squarely in scope.

What to watch

The immediate question is timing. The transparency obligations apply from August 2, 2026 — but the picture is complicated by the proposed AI Omnibus, on which the European Parliament and Council have reached political agreement. It contemplates targeted transitional relief for the Article 50(2) marking-and-detection duties, with the Council signalling a revised deadline of December 2, 2026 for those specific obligations. Formal adoption remains pending, leaving providers to plan against two possible dates.

The second is enforcement. Even non-binding guidelines, the Commission stresses, carry weight in how EU law is applied, and the CJEU retains the final word on interpretation. Whether a single icon can survive contact with the realities of live video, agentic systems and cross-border content will become clear only once the Code is signed and the rules go live. For now, in-scope organizations have a narrow window to audit their AI systems, disclosure flows and editorial workflows before the markings become mandatory.