Article 50 contains four different transparency paths

Article 50 of the EU AI Act does not create one universal “made with AI” label. It assigns different duties to providers and deployers for direct interaction, machine-readable marking, emotion or biometric systems, deepfakes, and certain text published on matters of public interest.

The provisions began applying on 2 August 2026, subject to the amended transitional rules. The Commission published final transparency guidelines in July 2026 and supports a voluntary Code of Practice for generated-content marking and labelling.

Use the exact paragraph and current guidance for each interaction and output path.

Direct interaction disclosure belongs at first contact

Article 50(1) requires providers of systems intended to interact directly with people to design them so the person is informed that they are interacting with AI, unless that fact is obvious to a reasonably informed, observant, and circumspect person in the context. Specific law-enforcement exceptions exist.

Evidence can include the interface name, welcome message, conversation header, voice introduction, onboarding, and accessibility behavior. Test a first-time user rather than relying on internal familiarity with the product.

Skybridge presents itself as an AI environment and labels agents in the interface. That is relevant evidence for a released interface. A future white-labelled chat embedded in another product would need its own assessment.

Provider-side marking is a technical output duty

Article 50(2) addresses providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video, or text. It requires outputs to be marked in a machine-readable format and detectable as artificially generated or manipulated, subject to technical feasibility, proportionality considerations in the text, and specified exceptions.

This is different from adding a visible footer to every email. The duty concerns provider-side technical marking. The Commission guidelines and Code describe the expected implementation in more detail.

An enterprise using an upstream model should ask whether the provider's released system supports the required marking and whether downstream processing preserves it. The answer belongs in the model and content route record.

Deployers have separate disclosure cases

Article 50(3) addresses deployers of specified emotion-recognition and biometric-categorization systems. Article 50(4) addresses deployers publishing deepfakes and AI-generated or manipulated text intended to inform the public on matters of public interest. Each path has its own exceptions and form of disclosure.

For public-interest text, the text includes an exception where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility. Do not extend that exception to unrelated paragraphs.

A routine supplier email drafted by an assistant is not automatically the same as public-interest publication. Skybridge's earlier internal tracker treated the missing “AI-assisted” footer as a general Article 50 gap. The current final guidance requires a narrower analysis of the actor, output, purpose, and paragraph. Document the conclusion for each outbound path in the release record.

Human review matters only where the provision says so

Human review can improve accuracy and accountability. It is not a universal transparency waiver. The reviewer also needs actual editorial control, material context, and responsibility where the legal exception requires them.

Record what the reviewer can see, change, refuse, and publish. A button shown after the content has already been distributed cannot supply editorial control.

The distinction fits Compsia's action design. A model may prepare a draft. Deterministic policy can stop the action. An authenticated person can approve the exact payload. These controls support the factual analysis while the law determines whether disclosure or marking is required.

What Skybridge must record for each released path

Create an Article 50 record containing:

Deepfake media sits outside the defined Skybridge production perimeter and triggers a separate capability and Article 50 review before release.

Test transparency as product behavior

Test a first-time user entering through every supported channel. Confirm the disclosure appears before or at first interaction and remains understandable with assistive technology.

For generated content, inspect the raw output and the final published artifact. Test whether editing, file conversion, copying, or a connector strips machine-readable marking. For editorial-review cases, verify that the named person can change or withhold publication and that the approved version matches the released content.

Store the result with the system version. A new provider, content modality, channel, white-label interface, or publication purpose triggers review.

Follow the content through every transformation. A provider may mark raw output before an application converts the file, adds a template, exports it, or passes it through a connector. Test the final artifact received by the audience. Record which party preserves the signal and which party supplies any visible disclosure. This makes transparency a supply-chain behavior with named ownership.

The wider EU AI Act compliance checklist shows where this Article 50 record belongs.

Article 50 questions

Must every AI-written email have a label?

Article 50 contains specific provider and deployer duties. A normal business email requires analysis of the exact system, actor, output, purpose, and applicable paragraph. Do not assume a universal footer rule.

Is calling a product an AI assistant enough disclosure?

It can be relevant evidence where the AI interaction is obvious in context. Test the actual first interaction and the Commission's current guidance.

Does human review remove all AI-content disclosure duties?

No. The public-interest-text provision includes a particular human-review and editorial-responsibility exception. Other duties have their own wording.