The transparency obligations that took effect on 2 August have moved from drafting into practice. Providers of the most advanced generative models must now ensure that users know they are dealing with an AI system, and that generated output is marked as such. In the weeks since, the first concrete implementations have been described, the first objection to them has been tested, and the Commission has answered a question practitioners have been asking for a year: where do AI agents sit in the Act?
How the marking actually works
Anthropic has described the method it will apply to future Claude models: a watermark that indicates the likelihood the model was involved in producing a given text. The company states that nothing is added to the text, that there are no hidden characters, and that a reader cannot distinguish marked output from unmarked output. Because no additional tokens are required, the measure carries no extra cost, and the mark itself holds no identifying information – it cannot be traced to a person, an organisation or a conversation. The practice will not be confined to one provider, since other developers that signed the same Code of Practice are to implement marks of their own.
The quality objection, and the answer to it
Watermarking works by influencing a model’s choice of words, which led critics to argue that it must by definition make writing worse. Euractiv reports the drafters’ response. Kalina Bontcheva of the University of Sheffield, who co-chaired the drafting of the transparency code, says the effect appears only in very short texts, and that the threshold was discussed with major providers and set at 200 tokens, roughly 150 words. Dino Pedreschi of the University of Pisa, her fellow co-chair, agrees, and a Commission spokesperson rejects the claim as well.
Reporting for POLITICO, Pieter Haeck sets out why the practical question is less about prose quality than about visibility. Ashley Casovan of the IAPP’s AI governance centre describes the change as a move from making an educated decision about whether something is AI-generated to seeing a verified mark. Walter Pasquarelli of the University of Cambridge notes that providers will have to work with platforms to display those marks, and points to evidence that engagement with material falls once audiences know it was generated.
Agents are not a new category
The Commission’s AI Act Service Desk has published guidance on AI agents, and its position is that no new category is needed. It accepts that the term is used inconsistently and is not clearly defined, while noting broad agreement that an agent must be able to take in information from its environment and act on it. On that basis the existing definitions – an AI system under Article 3(1), a general-purpose AI model under Article 3(63) – are treated as sufficient to capture agents already.
Three consequences follow. The prohibitions on harmful manipulation and on the exploitation of vulnerabilities may require safeguards at the design and development stage. The transparency rules apply from 2 August 2026 wherever an agent is intended to interact with people or to generate content. And an agent’s autonomy and its use of tools may be decisive in whether the underlying model is designated as carrying systemic risk, with the high-risk requirements following in December 2027 or August 2028 for systems classified that way.
What the taskforce discussed
The AI Office has also reported on the fourth meeting of the GPAI Signatory Taskforce, held on 17 July 2026, which took up the safety and copyright chapters of the Code of Practice. On safety, the Office explained how analysing the way a model is used can form part of post-market monitoring under Measure 3.5, complementing evaluations carried out before deployment. Participants discussed so-called marginal-risk clauses, under which a provider might match competitors that deploy unsafe models; the Office was clear that such a clause could be invoked only in exceptional circumstances and with evidentiary and procedural safeguards. On copyright, Measure 1.3(4) commits signatories to publishing which web crawlers they use and how those crawlers read robots.txt, and to notifying rightsholders of changes.
For anyone advising on deployment, the guidance on agents is the more consequential of the two developments. Watermarking is a compliance task with a known shape and a stated cost. The agent question decides which obligations attach to a system that plans and acts on its own, and the answer is that they attach already, under definitions written before the word became fashionable.
Source: The EU AI Act Newsletter #109: The Watermarking Era, August 24, 2026.