A rulebook that arrived two weeks late
On 20 July 2026 the European Commission published its final guidelines on Article 50 of the AI Act (Regulation (EU) 2024/1689) — the provision requiring AI providers and deployers to disclose chatbot interactions, machine-readably mark synthetic content, and label deepfakes. The guidelines arrived thirteen days before those obligations became legally binding on 2 August 2026, alongside a voluntary Code of Practice on Transparency of AI-Generated Content that the Commission had finalized separately on 10 June 2026 (nicfab.eu; European Commission).
The substance is not new — Article 50 has been law since the AI Act entered into force in 2024. What changed on 20 July was interpretive clarity: which systems count as "interactive," what counts as a machine-readable mark, and which uses are exempt. Getting that clarity two weeks before enforcement, for a rule companies have had two years to prepare for, tells its own story about how the Commission sequences AI Act implementation.
What the rule actually requires
Article 50 splits obligations by role. Providers must design AI systems so that a person interacting with a chatbot is told they're talking to a machine, unless that's already obvious. Providers must also mark AI-generated audio, image, video, and text as artificially generated, in a machine-readable and detectable format. Deployers carry a separate duty: deepfakes depicting real people or events must be disclosed as artificial or manipulated, and AI-generated or AI-manipulated text published on matters of public interest must be clearly labelled — unless a human has editorially reviewed it and someone holds editorial responsibility for its accuracy (artificialintelligenceact.eu).
The voluntary Code of Practice, drafted by six independent experts over a stakeholder process the AI Office launched in September 2025, gives companies a menu of watermarking, metadata, and provenance techniques to satisfy those duties and a presumption of compliance if they sign on (European Commission FAQ). Crucially, the Code itself concedes there is no single technical fix: "no single marking technology can currently meet all four requirements" of provenance, robustness, interoperability, and detectability (Tech Policy Press). Penalties for breaching Article 50 sit in the AI Act's second enforcement tier — up to €15 million or 3% of global annual turnover, whichever is higher (artificialintelligenceact.eu, Art. 99).
The case for the rule
The strongest argument for Article 50 isn't abstract. Synthetic audio and video have already been used to impersonate officials, fabricate evidence in disputes, and manufacture fake endorsements at a scale unmoderated markets can't self-police. A disclosure floor — telling people when they're talking to a bot, or when a video of a public figure is fabricated — is a modest, non-prescriptive intervention: it doesn't ban any model or use case, it just requires a label. Regulators are right that voluntary industry norms on watermarking have moved too slowly relative to how fast generation tools have proliferated, and a harmonized EU standard beats 27 divergent national ones.
Where the implementation falls short
But a rule engineered around good intentions still has to be implementable, and the sequencing here undercuts its own credibility. Publishing binding interpretive guidance thirteen days before enforcement gives providers effectively no lead time to align product design with the Commission's final reading of "machine-readable" or "obvious." The AI Omnibus proposal has already had to soften the blow by giving generative systems already on the market before 2 August until 2 December 2026 to meet the marking requirement — a tacit admission that the original timeline was unrealistic (artificialintelligenceact.eu).
The technical foundation is also thinner than the fine print of a €15 million penalty tier implies. The Commission's own Code acknowledges forensic detection tools remain unreliable and that no agreed evaluation standard exists for judging whether a marking technique actually works. Enforcing a hard legal duty against a technical baseline the regulator itself calls unsettled invites exactly the kind of inconsistent enforcement and defensive over-labelling that chills smaller developers — who lack in-house compliance teams — more than the large labs the rule is nominally aimed at. Open-source AI systems are explicitly not exempt, compounding that asymmetry.
None of this argues for scrapping disclosure requirements. It argues for what the Commission itself seems to have conceded with the Omnibus carve-out: phase mandatory marking in once measurable standards exist, keep the voluntary Code as the on-ramp rather than a compliance-in-name-only formality, and stop publishing decisive interpretive guidance in the final fortnight before a two-year-old law becomes enforceable. Transparency about AI content is a legitimate goal. Transparency about how the Commission itself is timing its own rulebook would help too.