Europe’s AI marking rules make the agency export a release decision

Provider marking, audience disclosure, provenance integrity and usage rights belong to separate checks at an agency handoff.

An agency can receive an AI-generated asset with provenance information and still lose track of what the audience will actually see. The decisive handoff is often mundane: a designer exports an image, an account team places it in an advertisement and another person approves the final crop. Europe’s current transparency framework makes it useful to distinguish the origin of the file from the disclosure accompanying the published communication.

This analysis was checked on 16 October 2026. It explains a bounded source distinction and proposes a fictional release exercise; it is not legal advice or a finding that any particular advertisement complies. No provider, advertising account or publication channel was connected.

Separate the provider’s marking from the deployer’s disclosure

The Commission [announced its marking and labelling code on 10 June 2026](https://digital-strategy.ec.europa.eu/en/news/commission-publishes-code-practice-marking-and-labelling-ai-generated-content); that announcement was last updated on 12 June. It describes a voluntary code supporting transparency obligations applicable from 2 August 2026. Voluntary participation in the code and applicability of the underlying obligations are different questions. Neither should be represented as an agency’s automatic certification.

The official [Article 50 text displayed by the AI Act Service Desk](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50) distinguishes provider machine-readable marking from deployer disclosure. It includes qualifications for assistive editing, deepfake content and evidently creative or fictional works. The public-interest text provision also has a human-review or editorial-control exception with editorial responsibility. These distinctions do not support a blanket assertion that every AI-assisted advertising sentence requires the same visible label.

The Commission’s [current code overview](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content) organises the work around marking and detection by providers and disclosure by deployers of covered deepfakes and certain text. An agency therefore needs to establish its role and the content’s characteristics before selecting an execution rule. A visible caption and a machine-readable signal serve different audiences and should not be treated as substitutes without examining the applicable requirement.

The [guidelines landing page](https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems) records publication on 20 July 2026 and a later update on 31 July. That is a separate chronology from the June code announcement. The linked full guidelines file was not available through this research route, so this article does not claim a paragraph-by-paragraph review of that document. Particular legal decisions need the authoritative materials and qualified review.

Provenance answers a narrower question than truth

The [C2PA version 2.2 explainer](https://spec.c2pa.org/specifications/specifications/2.2/explainer/Explainer.html) describes signed provenance information bound to content. It explicitly separates verification of that information from judging whether its assertions are true. A valid provenance record is therefore useful evidence about a file’s history, not proof that a product claim is accurate, a depicted person consented or a licence covers a campaign. This is a version-specific reference, not a claim that version 2.2 is the newest specification.

A practical release record should keep those questions apart. Record the source asset, the permitted use, the described transformation and the intended audience-facing disclosure. Assign someone to verify product facts independently of the provenance check. When the input is a supplier photograph, the record should identify the permission actually granted; the mere presence of a signed content credential should not be used to fill a missing permission field.

Inspect the final export as a distinct deliverable. The proposed check is simple: does the exported file match the approved creative, and does the intended placement retain the disclosure in a readable location? This article does not report that a named platform strips metadata or captions. It recommends testing the actual route instead of assuming that information attached to a working file necessarily reaches the audience unchanged.

Keep the review understandable to people outside production. A client should be able to see which element is generated, what remains a real product photograph and what was checked before release. Avoid a generic declaration that an entire asset is trustworthy. A precise description of the transformation gives a reviewer something concrete to compare against the actual image and its intended use.

Fictional exercise: a lamp in an imagined room

Consider a fictional Lisbon agency preparing an advertisement for a real, photographed table lamp in an invented interior. This is an unrun learning exercise, not an actual client commission. The agency would need permission to use the lamp photograph and any other inputs before creating a real asset. The imagined room must not be presented as an installed project or a customer’s home.

The proposed asset sheet distinguishes the supplied product photograph from the generated surroundings. A reviewer compares the lamp’s dimensions, fittings and visible materials with the approved product record. If a variation changes the shade or creates an unsupported lighting effect, it returns for correction. A disclosure about the room’s origin would not resolve an inaccurate depiction of the product being sold.

The exercise then follows the asset through a wide advertisement, a narrow phone placement and an exported presentation. For each, the reviewer checks the complete composition and accompanying language. The desired statement identifies the interior as an illustration rather than implying a completed installation. The precise disclosure and legal treatment remain decisions for the responsible team based on the actual final content and its context.

Next, the designer examines the exported file’s available provenance information and records the observation without turning absence into proof of deception or presence into proof of truth. The exercise deliberately allows an unresolved outcome. If the workflow cannot preserve or verify the required signal, the team should hold that route for review rather than invent a successful verification receipt.

Finally, the account lead signs a release record referring to the exact export, permitted destination and approved disclosure. The record is not a legal compliance badge. It is an operational acknowledgement of what was reviewed. If a later editor changes the room, crop or accompanying claim, the earlier acknowledgement remains historical evidence and should not silently approve the changed communication.

Make uncertainty visible before distribution

The most useful failure state is specific and recoverable: the file is ready for visual review, but the permission record or required disclosure is unresolved. Preserve the creative and explain the missing decision. Avoid announcing a generic technical failure to the client or marking the entire campaign compliant because one automated check succeeded. Human review needs both a usable artefact and an honest account of the remaining gap.

An agency can also build a comparison exercise without generating new content. Take an authorised example asset, describe the proposed transformations and ask reviewers to classify which claims about its origin and product remain supportable. Disagreement reveals where the production brief is ambiguous. It is a reason to clarify the brief, not evidence that a legal exception necessarily applies.

The wider reporting lesson is that marking, disclosure, factual accuracy and usage rights are connected but separate controls. The current European materials make those boundaries explicit enough to improve a release workflow. They do not supply a universal caption or remove the need to examine the actual advertisement. The final export, its explanation and its permissions should arrive at approval together.