EU AI Act Article 50:
an agency checklist.

From 2 August 2026, Article 50 transparency obligations apply. For agencies, the hard part is not adding “made with AI” to everything. It is identifying which role and content path applies, making a proportionate disclosure when needed, and preserving the facts behind the publication decision.

1. Start with the right role

The European Commission guidance separates providers, who develop or place an AI system on the market under their name, from deployers, who use a system under their authority for professional purposes. An agency using an existing model to create client work will often be operating as a deployer, but custom systems and branded tools can change the analysis.

The obligations can also overlap. A system can be interactive, generate marked output, and produce content that requires a visible deployer disclosure. Start by recording the system, the intended use, the final content type, and the organisation responsible for publication.

2. Interactive AI: disclose at first interaction

Providers of AI systems designed to interact directly with natural persons must generally make the artificial nature of the interaction clear. The Commission gives a first-turn chatbot statement—such as “You are interacting with an AI system”—as a straightforward example.

Do not rely only on terms and conditions, a buried URL, or a generic note that a website “uses AI.” The guidance says the information should be clear, distinguishable, and available at the first interaction. Agencies delivering chatbots or voice agents should record the exact wording, placement, language, and a dated screenshot or transcript showing that it appeared.

3. Realistic image, audio, or video: test the deep-fake path

The guidance describes a deep fake as AI-generated or manipulated image, audio, or video that appreciably resembles a real or plausibly real person, object, place, entity, or event and could falsely appear authentic or truthful.

Not every AI-assisted visual is automatically a deep fake. Minor technical adjustments and clearly impossible or fictional scenes may fall outside that definition. Context matters. The Commission specifically contrasts an AI-generated, realistic product image that could mislead viewers with a real product shown against a synthetic background that does not misrepresent the product.

When the final asset may qualify, the deployer disclosure needs to be understandable and perceivable without special tools. Provider metadata alone is not the same as a visible or audible audience disclosure.

4. Public-interest text: record review and responsibility

AI-generated or manipulated text published to inform the public about matters of public interest has a separate disclosure path. The regulation includes an exception where the publication has undergone human review or editorial control and a person or organisation holds editorial responsibility.

A vague claim that “someone looked at it” is poor evidence. Record the reviewer, review date, what was checked, what changed, and who controlled and accepted responsibility for the publication.

5. Build a publication evidence record

The Commission says organisations outside the voluntary Code of Practice should be able to explain how their alternative measures meet the obligations. A compact record helps the client, account team, and future reviewer reconstruct the decision without pretending to be a legal certification.

• Client, campaign, asset, channel, and publication date
• AI system, model version, and the contribution it made
• A fingerprint of the final file or an exact final URL
• Human reviewer, review date, and editorial owner
• Whether provider markings were retained
• Exact disclosure wording, placement, and timing
• Rationale, exceptions relied on, and supporting links

6. The five-minute pre-publish checklist

1. 01
   Identify the output.

   Text, image, audio, video, interactive AI, or a mixed experience?

2. 02
   Describe the AI contribution.

   No AI, assistive edit, substantive generation, or manipulation?

3. 03
   Test the narrow disclosure paths.

   Interactive AI, potentially deceptive realistic media, or public-interest text?

4. 04
   Complete and record human review.

   Name the reviewer and responsible publisher; do not record an intended review as completed.

5. 05
   Preserve the final facts.

   Fingerprint the asset, copy the exact label, note its placement, and export the record.

Primary sources

European Commission: Guidelines on transparency obligations for providers and deployers of AI systems, published 20 July 2026.

European Commission: Code of Practice on Transparency of AI-generated Content, including provider marking and deployer labelling measures.