Article 50 Machine-Readable Marking for AI-Generated Text: Scoping AI-Drafted Customer Letters

What Article 50(2) actually requires

Article 50 sits in Chapter IV of Regulation (EU) 2024/1689, titled transparency obligations for providers and deployers of certain AI systems. Paragraph 2 provides that providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Text is named expressly.

Two points matter for scoping. First, the duty is addressed to providers. In its wording it is not addressed to deployers. Second, it is a design and output obligation on the system. It is not a drafting rule for each document an organisation produces with that system.

The paragraph also sets a standard for the technical solution. It must be effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities and limitations of various content types, implementation costs and the generally acknowledged state of the art. A recital lists watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity, logging methods and fingerprints as possible techniques, as may be appropriate. It does not prescribe one for text. Compliance teams should not assume a single marking method exists or that any given method will survive editing.

Provider and deployer roles in a multi-tool drafting chain

A common assumption is that any organisation sending AI-drafted text must mark it. The wording of Article 50(2) does not support that reading. The first question is a role question: for each AI system in the letter workflow, is the enterprise the provider or the deployer? The roles are defined in Article 3 and should be confirmed there for each system.

In a typical workflow the enterprise calls a third-party general-purpose model, perhaps through an intermediary tool, then applies templates, edits and approvals before the letter is sent. Each of those components may have a different provider. An enterprise that only uses these systems under its own authority is generally positioned as a deployer. One that builds its own drafting system on top of a model, or offers it under its own name, needs a closer analysis of whether it has taken on provider responsibilities.

This has a practical consequence. Marking is something to ask of the vendor, in contracts and documentation: does the system mark text outputs in a machine-readable format, and does that marking persist once the enterprise edits, reformats or templates the draft? Any marking that does not survive those steps should be treated as a limit on what can be relied on downstream, not as a control the enterprise owns.

Article 50(4), public-interest text and human review

Deployer duties sit in Article 50(4). It requires deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake to disclose that the content has been artificially generated or manipulated. For text, the paragraph is read as applying to text published to inform the public on matters of public interest, with an exception where the text has undergone human review or editorial control and a person holds editorial responsibility. Counsel should confirm the exact wording of that limb and exception against the current consolidated text before the scoping decision is finalised.

Applied to customer letters, the test has two parts. The first is whether the letter is text published to inform the public on matters of public interest. A letter addressed to one customer about that customer's account is not self-evidently a publication to the public. Standardised notices sent to large populations, or text that is also posted publicly, may deserve a closer look. The second is whether human review and editorial responsibility applies, and that depends on what the review actually is. A named person who reads the draft, can change it and accepts responsibility for it is different from a nominal click-through.

This is a judgement for the enterprise and its legal advisers. Treat it as a documented position with a review date, because guidance and codes of practice on marking and labelling may develop. The exception only helps if the review it relies on can be shown to have happened.

Build a scoping register by letter category

A scoping register turns the legal reading into something operations can apply. Keep it at the level of letter category, such as notices, correspondence and explanations, because the analysis can differ between them.

• Letter category and channel: What the letter is, who receives it, and whether any version is published beyond the individual recipient.
• Systems in the drafting chain: Each model and tool that touches the draft, with the party that is provider and the enterprise's role.
• Paragraphs considered: Article 50(2) and Article 50(4), with the reasoning for why each does or does not apply to the enterprise.
• Review step relied on: The human review required before sending, who may approve, and what authority they hold over the text.
• Vendor marking position: Whether the provider marks text outputs and whether that marking persists through editing and templating.
• Owner, sign-off and review date: Named legal and compliance owners, the date of approval and the date the position will be revisited.

Where runtime enforcement fits

Trussed enforces policy in the execution path across models, agents, tools and MCP servers. Decisions are made on individual interactions and tool calls: allow, redact, block, rate limit or escalate for human approval. For AI-drafted letters, an enterprise can express its scoping decision as policy, for example requiring human approval before a given category of draft is released. That policy is enforced at runtime, not reviewed after the fact.

The enforcement generates audit-ready evidence continuously, mapped to the EU AI Act among other frameworks. Trussed can be deployed self-managed in the customer cloud, with configurable retention and data residency.

The limits are as follows. Trussed does not apply watermarks or mark content, so it does not satisfy a provider's Article 50(2) duty. It does not make legal determinations about scope, and it does not provide model provenance or end-to-end tracing of agentic workflows. It evidences what was enforced on each interaction under the policy the enterprise defined.