Anthropic is a signatory to Section 1 of the EU’s Code of Practice on Transparency of AI-generated Content — the provider section, addressing Article 50(2) and (5).
Section 2 of the same Code addresses deployers under Article 50(4) and (5), and it is a separate signature. Our observation is that these are two obligations at two boundaries, and the upstream one does not discharge the downstream one. What connects them is reconciliation, not inheritance.
Two sections, two signature lists
The Code has two sections and the Commission publishes the signatories separately: 82 organizations to Section 1, 152 to Section 2, roughly 190 in total. Anthropic appears under Section 1, alongside Google, Meta, Microsoft, Mistral, OpenAI, and Cohere. Section 2’s list is populated by deploying organizations across industry — Getty Images, Lenovo, Lufthansa, Iberdrola among them.
Section 2 applies to a signatory only insofar as it is a deployer of AI systems that generate or manipulate image, audio, or video content constituting a deep fake, or text published to inform the public on matters of public interest. It is scoped to a use, not to an industry.
Two structural properties of the Code matter before any of its measures do. Adherence is voluntary while the Article 50 obligations it supports are binding, and the Commission and the AI Board have confirmed the Code adequate for demonstrating compliance. And Section 2 states in its own objectives that adherence does not constitute conclusive evidence of compliance with the obligations under the AI Act. A signature is a recognized route, not a determination.
What Article 50(4) asks of a deployer
The obligation runs to disclosure, and the Code reproduces the legal text.
A deployer of a system that generates or manipulates content constituting a deep fake discloses that the content has been artificially generated or manipulated. A deployer of a system that generates or manipulates text published to inform the public on matters of public interest discloses the same. Article 50(5) sets the manner: clear and distinguishable, at the latest at the time of first interaction or exposure, conforming to applicable accessibility requirements.
Three exceptions sit in the text. Use authorized by law to detect, prevent, investigate, or prosecute criminal offenses. Content forming part of an evidently artistic, creative, satirical, fictional, or analogous work, where the obligation narrows to disclosing the existence of the generated content in a manner that does not hamper the display or enjoyment of the work. And for published text, where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication.
The Commission’s Article 50 guidelines, which complement the Code, define the concepts these exceptions turn on — including AI-generated text on matters of public interest — and set out practical examples of what falls in and out of scope; the guidelines were adopted July 20, 2026.
None of these exceptions turns on whether the model that produced the content was marked.
The human-review exception is a structure
The third exception is the one most likely to be read as an escape route, and Section 2’s Commitment 4 sets out what relying on it involves.
Media service providers within the meaning of Regulation (EU) 2024/1083 may rely on existing review and editorial procedures and established professional standards. Every other signatory commits to establish, adapt, or maintain policies for human review or editorial control prior to publication, proportionate to size and resources, including at minimum the identification of the natural or legal person holding editorial responsibility — name, role, and contact details — and an overview of the concrete organizational measures and human resources allocated to ensure adequate review is performed and responsibility assumed before publication. Where not already public, those contact details are published.
The Code is explicit that this does not extend to documenting individual instances of review over individual publications, and that implementation must not affect media freedom, editorial independence, or protection of journalistic sources.
Stated precisely: the exception is available on the condition that a named accountable party and a resourced process exist and are visible. That is an assurance structure with an owner, which is a different thing from an exemption.
Disclosure carries its own persistence problem
The companion dispatch traced a machine-readable mark through enterprise transforms. Section 2 puts the same problem one layer up, at the human-perceptible label.
Measure 1.2 requires the icon or equivalent label to be placed so that it ensures immediate recognition without requiring user interaction or sustained attention; to remain visible long enough to be noticed under normal exposure conditions; and to be directly embedded into the content unless equivalent alternatives are available, such as an interface overlay that appears to the viewer to be on the content. Signatories are further encouraged to collaborate on a best-effort basis with the actors whose services they use to distribute the content — publishers, online platforms, retail — to preserve the label so that disclosure accompanies the content throughout its distribution and dissemination chain, online and offline.
Best-effort collaboration across a distribution chain is not a control that binds. So the label faces the same question the mark does: it is applied at one point and must survive a path the applying party does not fully own. Where a machine-readable mark is at risk from re-encoding and format conversion, a perceptible label is at risk from cropping, re-hosting, recomposition into a new surface, and platform interfaces that strip or reposition overlays.
Where visual disclosure is not possible, the Code specifies an audible disclaimer at the beginning of the content in plain language, with an earcon or other audible solution permitted pending a common EU-wide approach from the Task Force. Accessibility obligations run alongside, referencing the European Accessibility Act, the Web Accessibility Directive, ETSI EN 301 549, and WCAG 2.1.
Where the verification gap actually bites
Section 1’s Commitment 2 requires providers to make a detection solution available free of charge, in a form such as a public specification, software, or a cloud API, and to make it openly accessible where the general public may be exposed to the content. Free-form text is the stated exception: a provider may restrict access to text-watermark detection to verified expert users with a legitimate need, on grounds of lower reliability and robustness and the risk of misleading or low-confidence results — and any such restriction is limited in time, until more reliable mechanisms emerge as state of the art. Anthropic’s detection documentation had not been published as of this writing.
Worth drawing the boundary carefully. A deployer’s Article 50(4) duty attaches to content it generates or manipulates, so it does not need a detector to know disclosure is owed. The gap is not in knowing; it is in demonstrating. Verifying that a supplier’s or contractor’s delivered content carries what it should. Auditing a pipeline for content that was published without a label. Responding to a substantiated report of missing or incorrect disclosure, which Measure 2.3 of Section 2 commits signatories to review and remedy without undue delay. Each of those is an evidence task, and each may run against a detection route that is permitted to be gated.
Reconciliation, not inheritance
The structure is familiar from regulated data handling. Under a business associate arrangement, two parties are each accountable at their own boundary; neither inherits the other’s compliance, and the connective tissue is contractual reconciliation rather than one party preventing the other’s failure. Article 50 divides marking and disclosure the same way. The provider marks and makes detection available at its boundary. The deployer discloses at its own. A Section 1 signature upstream is an input to the deployer’s posture and not a discharge of it.
The dates compress this. Article 50 applies from August 2, 2026. Generative systems already on the EU market before that date have until December 2, 2026 for Article 50(2) marking; systems placed on the market on or after August 2 get no such period. Signatory Taskforces stand up in September 2026, and Section 1’s watermark-detection interoperability requirement carries a deadline of February 2, 2027. The provider side of the seam is still being built while the deployer side is already in force.
The hard claim
A provider’s signature covers the provider’s boundary. Article 50(4) creates an obligation at the deployer’s boundary that no upstream commitment satisfies, that does not depend on whether the upstream model was marked, and that the Code itself declines to treat as conclusively discharged by signature.
What the enterprise inherits from a provider is evidence it can cite, not a duty it can transfer.
Which makes the deployer’s work structural rather than declarative: knowing which of its uses fall inside Article 50(4), whether it relies on the human-review exception and can name the accountable party, how its labels survive the distribution chain it does not own, and how it would demonstrate any of this if the route to verifying the underlying mark is gated.
Which makes the deployer’s work structural rather than declarative: knowing which of its uses fall inside Article 50(4), whether it relies on the human-review exception and can name the accountable party, how its labels survive the distribution chain it does not own, and how it would demonstrate any of this if the route to verifying the underlying mark is gated. That is a design problem at the disclosure boundary, and it belongs to the enterprise deploying the system.
