Does the AI Act Code of Practice on labelling AI-generated content strike the right balance for 2 August 2026?

► President
I open this deliberation on the Code of Practice for marking and labelling AI-generated content, published under Article 50 of the AI Act. The transparency obligations it accompanies apply from 2 August 2026. The question is narrow but consequential: does this Code strike the right balance between giving providers and deployers legal clarity and delivering the protection against deceptive synthetic media that citizens expect? The legal basis is settled — Article 114 TFEU, the same internal-market basis as the AI Act. This is a soft-law implementation instrument, not new legislation, so it needs no legislative adoption — but it carries the Commission's name. I flag three tensions in advance: the free-expression boundary, the extraterritorial reach onto non-EU providers, and the compliance burden on SMEs and open-source models.
[EUR-Lex — verify Art. 50 / Art. 114 Regulation (EU) 2024/1689]

► EVP · Digital
This portfolio leads. Article 50 sets transparency obligations — the AI Act's limited-risk tier, not the high-risk regime. Four pillars, mirrored in the tweet: chatbot disclosure unless obvious from context; disclosure of deep fakes and AI text on matters of public interest; machine-readable marking of synthetic audio, video, images and text to the extent technically feasible; and the Code itself translating these into watermarking, metadata and provenance standards. My recommendation is to note and endorse the framework, but I concede two soft spots. First, the phrase to the extent technically feasible is doing heavy lifting — watermark robustness against cropping, re-encoding and adversarial removal is not solved, and the Code must not promise detection it cannot deliver. Second, adherence is voluntary; non-signatories still owe the Article 50 obligation but lack the safe harbour. This is the AI Office's first real transparency-enforcement test.
[EUR-Lex — verify Art. 50 Regulation (EU) 2024/1689]

► EVP · Democracy
Strong support, with a caveat from the other direction. Synthetic media on public-interest matters is the core disinformation threat to democratic processes, and labelling deep fakes is exactly the European Democracy Shield agenda. My concern is under-reach, not over-reach. The human editorial review carve-out for AI text must not become a loophole that lets coordinated influence operations launder synthetic text through a thin editorial veneer. The Code must define that carve-out tightly. Charter Article 11 cuts in favour of transparency here — citizens have a right to know.
[EUR-Lex — verify Art. 11 Charter; Art. 50(4) carve-out]

► EVP · Economy
Reservations. I run the competitiveness lens. The Article 50 obligation falls hardest on the smaller end — GPAI start-ups and SME deployers who must retrofit marking and provenance pipelines. The Code must not become de facto mandatory tooling that only large incumbents can afford, which would entrench the very gatekeepers we are trying to discipline elsewhere. I want the proportionality clause and the SME guidance made explicit, consistent with the Competitiveness Compass and the omnibus simplification line. Endorse only with that condition.

► Commissioner · Justice
Reservations. The Charter is a legal limit, not a slogan. Article 50 labelling engages Article 11 freedom of expression for satire, parody and artistic work, and Article 11 also protects the right to receive information, which supports disclosure. The proportionality test is met for deep fakes on public-interest matters. It is not automatically met for all creative or clearly fictional synthetic content. The Code must preserve the AI Act's own exemption for evidently artistic, satirical or fictional works, carried through verbatim from Article 50(4), or it will not survive a Charter challenge.
[CJEU — verify Art. 11 Charter proportionality doctrine]

► Commissioner · Trade
Reservations on extraterritorial reach. Article 50 binds any provider placing a GPAI system on the EU market, including non-EU providers. A marking standard that effectively prescribes a third-country provider's product design engages WTO-TBT Article 2.2 — not more trade-restrictive than necessary — and is a live topic at the EU-US Trade and Technology Council. If the Code points to one proprietary watermarking standard, that is a technical barrier to trade and an actionable grievance. My condition: the Code must stay standard-neutral and reference open international standards rather than mandating a single vendor's method.
[EUR-Lex — verify WTO-TBT Art. 2.2 applicability]

► Commissioner · Internal Market
Support, with an industrial caveat. A single EU-wide Code is exactly the single-market logic — 27 divergent national labelling regimes would fragment the market and is the worse outcome. But I echo Economy on SMEs and add the open-source dimension. Free and open-source GPAI models released under the AI Act's open-source provisions need a clear, lighter path, or the Code chills the European open-model ecosystem we depend on for sovereignty. Standardisation should be the lever — make the marking standard a public European standard, not a compliance tax.

► Commissioner · Competition
Reservations on asymmetry, but no red line. My concern is that a voluntary Code plus a safe harbour rewards scale: gatekeepers can fund provenance infrastructure and signal compliance, while smaller rivals cannot. That is a competition distortion dressed as transparency. I do not oppose the Code, but I want monitoring of whether adherence becomes a competitive moat.

► Commissioner · Education & Culture
Support, with a cultural note aligned to Justice. AI is now a legitimate creative tool, and over-labelling chills cultural production. The artistic, satirical and fictional-works exemption is essential. I endorse provided that exemption is explicit and that disclosure for creative works is appropriate and non-intrusive, in line with Article 50(4).

► Commissioner · Research & Innovation
Reservations. The research-model dimension matters here. Academic and research GPAI releases must not be caught as if they were commercial deployments. I support the Code if the research and open-source exemption tracks the AI Act's own wording rather than going beyond it.

► HR/VP · Foreign Policy
Support. Synthetic media is a primary tool of foreign information manipulation and interference. A robust EU labelling baseline strengthens our hand in the G7 and with like-minded partners on AI provenance norms. I align with Trade that the standard must be internationally interoperable, not unilateral.

► President
I read the room. There is a clear majority — no portfolio opposes the Code. I name the tensions. Justice and Education-Culture want the artistic and satire exemption protected against over-reach, while Democracy wants the editorial-review carve-out for AI text tightened against under-reach; the Code must hold both lines at once. Trade and Digital diverge on extraterritorial reach and WTO-TBT exposure unless the marking standard stays vendor-neutral and internationally interoperable. Economy, Internal Market and Research want the SME, open-source and research burden contained and the feasibility clause made honest. There is no blocking objection — all reservations are conditions. This is not a legislative adoption; Article 50 applies on 2 August 2026 regardless. I record consensus to endorse the application framework with conditions, and I will not call a formal vote.

Deliberation closed · 12 positions recorded