EU AI Act Article 50 transparency requirements: practical checklist

Who does Article 50 apply to?

Assess the provider and deployer roles separately. Developing or commissioning an AI system under your own name may make you a provider; using a system under your authority may make you a deployer. Being based in Switzerland or the US does not by itself settle scope: review the EU market, users and where outputs are used. One organization can have different roles for different systems.

Official source for this requirement

• Operational starting point: create one record per system and use case, not one yes/no statement for the whole company.
• Record the facts behind the role assessment and assign uncertain cases to a qualified reviewer.

1. AI chatbot disclosure: direct interaction

Article 50(1) addresses providers of AI systems designed to interact directly with people. People must be informed that they are interacting with AI unless this is obvious in the circumstances. Article 50(5) requires relevant information to be clear, distinguishable and accessible, at the latest on first interaction or exposure. Do not assume a company privacy-policy link is an adequate chatbot notice.

Official source for this requirement

• Illustrative wording for review: “You are chatting with an AI assistant.” Add a human-support route only if one actually exists.
• Our suggested acceptance check: open a fresh session on mobile and desktop. Capture what appears before the first message, including the language and any screen-reader label.
• For a third-party widget, ask who controls the notice, whether you can remove it accidentally and what happens when the supplier updates the interface.

2. Synthetic-content marking: technical evidence

Providers of systems generating synthetic text, images, audio or video face machine-readable marking and detectability requirements. The Commission distinguishes this from visible labels for audiences and describes exceptions for standard editing or changes that do not substantially alter the input or its meaning. Obtain evidence about the actual system and output formats rather than treating a watermark icon as proof.

Official source for this requirement

• Suggested vendor request: identify the marking method, supported exports, detection method, limitations and release version.
• Suggested workflow test: save an original export, then process it through your normal publishing tools. Compare what evidence survives resizing, compression or format conversion.
• If a test fails, record the affected channel and ask the supplier and reviewer for a resolution. Do not label an untested method effective merely because it is documented.

3. Deepfake disclosure: what the audience sees

Deployers must disclose AI-generated or manipulated image, audio or video that constitutes a deepfake. Context matters: assess whether the content misleadingly appears authentic, not merely whether AI was used. The Commission explains that machine-readable marks alone do not meet the human-facing disclosure obligation. Evidently artistic, creative, satirical or fictional works have a tailored disclosure approach, not an automatic blanket exemption.

Official source for this requirement

• Illustrative scenario: a synthetic spokesperson appears to deliver a real customer testimonial. Record the represented person, claims, audience and publication context before deciding how to proceed.
• Illustrative wording for review: “This video contains AI-generated imagery and voice.” Wording must accurately describe the specific material.
• Suggested check: review the placement in the actual feed or player, including muted playback and cropped previews. Store the final published version with the approved wording.

4. AI-generated public-interest text and editorial review

Article 50(4) addresses AI-generated or manipulated text published to inform the public on matters of public interest. Its exception requires human review or editorial control and an identifiable natural or legal person holding editorial responsibility. This is not a rule that every AI-assisted internal draft needs a public label.

Official source for this requirement

• Our suggested review record: draft identifier, factual sources checked, substantive changes, reviewer, approval date and responsible publisher.
• Keep the final version alongside the reviewed draft. A later automated rewrite should trigger a fresh publishing check.
• Use a documented review workflow instead of a checkbox that silently defaults to “human approved”.

5. Emotion recognition and biometric categorisation

Article 50(3) requires deployers of these systems to inform exposed people and comply with applicable personal-data rules. A transparency notice does not make an otherwise prohibited use permissible. Escalate the use case before treating disclosure as the only action needed.

Official source for this requirement

• Our suggested intake record: purpose, setting, affected people, data categories, supplier, decision use and responsible owner.
• Keep the decision to allow or stop the system separate from approval of the notice wording. This prevents a communications task from becoming accidental authorization to deploy.

Dates: avoid assuming a blanket grace period

Article 50 applies from 2 August 2026. The Commission currently describes a limited transition to 2 December 2026 for the Article 50(2) marking and detection obligation for systems placed on the market before 2 August. This is not a general extension for chatbot notices or deployer disclosures. Confirm the current legal position and the relevant system date before relying on the transition.

Official source for this requirement

• Record the source consulted, date checked, system release evidence and reviewer. Revisit the record when the system, guidance or intended use changes.

Worked example: one campaign, three separate checks

Fictional example: a small retailer launches a support chatbot, a synthetic product-demo video and an AI-assisted article. The operations owner creates three records instead of marking the whole campaign “Article 50 complete”. The chatbot record captures the first-interaction notice and vendor responsibilities. The video record asks whether it constitutes a deepfake and records the relevant disclosure decision. The article record captures its purpose, editorial review and responsible publisher. These are proposed operating records, not legal conclusions.

• Status choices: not assessed, evidence requested, changes needed, ready for qualified review, reviewed with conditions.
• Each unresolved item gets one owner and one next action. For example: “Marketing owner to capture the mobile video preview with the proposed disclosure before publication.”
• Completion means evidence and a reasoned decision exist. It does not mean a template or checker has certified compliance.

Official sources

• EU AI Act — Article 50 and scope provisions (EUR-Lex)
• European Commission — Article 50 questions and answers
• European Commission — quick facts on transparency rules
• European Commission — transparency guidelines