Anthropic has drawn attention for putting machine-readable watermarks into text produced by newer Claude models. But for agencies, the more consequential part of the EU AI Act may sit one paragraph further down the law.
Article 50(4) creates a separate transparency duty for deployers of generative AI systems that publish AI-generated or manipulated text to inform the public on matters of public interest. The rule also contains an exemption when the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility. For agencies using AI in client communications, that turns a familiar production step into something with regulatory weight.
Table of contents
Jump to each section:
- What Article 50(4) actually changes for agencies
- Human review is doing more work than agencies may think
- Claude’s watermark is evidence of processing, not authorship
- The operational question is who takes editorial responsibility
- Why this becomes an agency workflow issue now
Key Takeaways
- Article 50(4) can require deployers to disclose AI-generated or manipulated public-interest text.
- Substantive human review or editorial control can remove that labelling requirement when editorial responsibility is also held.
- A Claude watermark signals processing by Claude, not proof that Claude authored the underlying text.
What Article 50(4) actually changes for agencies
The distinction starts with who the law regulates. Article 50(2) is aimed at providers of generative AI systems, requiring machine-readable marking of synthetic outputs. Article 50(4), by contrast, reaches deployers using generative AI systems in professional contexts when they publish certain content.
The EU AI Act text says the disclosure obligation covers AI-generated or manipulated text published to inform the public on matters of public interest. The Commission’s later guidance says that can include economic, financial, political, scientific and cultural developments that may be relevant to public debate.
That does not mean every AI-assisted agency deliverable needs a label. A product description, internal memo or sales email may fall outside this specific public-interest test. But public-facing thought leadership, policy commentary, corporate statements on consequential events and some forms of editorial content can sit much closer to it.
The practical issue is therefore not simply whether an agency used AI. It is whether the agency is acting as a deployer, what kind of text is being published, what purpose the publication serves, and what human editorial process sits between generation and release.
Human review is doing more work than agencies may think
The exemption is where agency workflow becomes important. The European Commission’s Article 50 guidance says published text that has undergone human review or editorial control does not need the Article 50(4) label when editorial responsibility is held.
But the Commission draws a line between substantive review and cosmetic checking. Human review means deliberate examination of the substance by people with relevant knowledge and professional judgment. Editorial control means an editorial entity has authority to approve, alter or reject the substance, including fact-checking and assessing the trustworthiness of sources.
The Commission explicitly says superficial checks such as spell-checking or grammatical correction do not count. That is a meaningful distinction for agencies that currently describe their workflow as “human reviewed” when the human step is mostly cleanup before publication.
A workflow in which an account executive scans a draft for tone and typos is not the same thing as an editor checking claims, sources, framing and accuracy before taking responsibility for publication. Agencies that want to rely on the exemption may need to be able to explain which of those processes they actually run.
Claude’s watermark is evidence of processing, not authorship
Anthropic’s own marking policy shows why disclosure and watermarking should not be treated as the same question. Its documentation says Claude models launched on or after 2 August 2026 support marking at launch, with marks applied across supported Claude surfaces worldwide, including the API, Claude Code, Claude Cowork and Claude Tag.
More importantly, Anthropic characterizes a detected mark as a signal that content may have been “processed by Claude.” Ahrefs highlighted that processed-versus-authored distinction because it matters operationally. A writer can draft a piece themselves, then use Claude for proofreading, translation or summarization, and the returned output can still contain Claude-influenced wording.
That makes the watermark poor evidence for deciding whether a publisher complied with Article 50(4). The legal question is about the nature of the content and the deployer’s review and editorial responsibility. The watermark is a provenance signal showing model involvement somewhere in the pipeline.
Conflating the two could create a new category of agency-client disputes. A client may see a detected mark and assume the agency outsourced authorship to AI, while the agency may have used the model only to edit human-written copy. The mark cannot resolve that disagreement by itself.
The operational question is who takes editorial responsibility
The primary EU text also corrects a detail that is easy to lose in secondary summaries. The exemption applies when a “natural or legal person” holds editorial responsibility, not only when a named individual personally assumes it.
For agencies, that means process design should focus on clear accountability rather than a ritual sign-off. Someone with real authority needs to own the publication decision, and the review preceding that decision needs to engage with substance.
That could affect how agencies structure AI-assisted work for clients. If the client is the publisher and retains final editorial control, the client’s process may be central. If the agency operates a newsroom, publishes directly on a client’s behalf or controls the final text, the agency may have a different level of deployer responsibility. Those distinctions are fact-specific and should not be collapsed into a blanket rule that “agencies must disclose AI.”
What agencies can do now is document the chain: where AI enters, what it changes, who checks the factual substance, who can reject or rewrite it, and who holds final editorial responsibility.
Why this becomes an agency workflow issue now
Article 50’s transparency obligations have applied since 2 August 2026. The European Commission says about 190 organisations signed its Code of Practice on Transparency of AI-generated Content by the end of July, with major providers including Anthropic, Google, Meta, Microsoft, Mistral and OpenAI represented.
So Claude is the visible example, not the whole story. The provider side of the market is moving toward more machine-readable provenance. At the same time, the deployer side is being asked to make its own disclosure and editorial-control decisions.
For agencies, the immediate question is less “Can a client detect that we used AI?” and more “Can we show what meaningful human editorial control looked like before this went live?” That is a process question, a client-governance question and, for some public-interest content, now a legal one too.
