EU AI Act Article 50: plan transparency duties in practice

The short answer on Article 50

Article 50 of the EU AI Act has applied since 2 August 2026. It creates different transparency duties for providers and deployers of certain AI systems: informing people about AI interaction, marking synthetic outputs in a machine-readable form, and disclosing deepfakes and certain AI-generated public-interest text. The applicable duty depends on the organisation’s role, system and use case. This guide supports planning and is not legal advice.

One label therefore does not solve every requirement. A technical provenance signal embedded in content and a clear notice for people serve different purposes. Organisations should first classify their role and the specific publication before defining templates, approvals or training.

Classify the role, system and use case separately

For each use case, determine whether your organisation provides an AI system under its own name or uses it under its authority. The same organisation can be a provider, deployer or both, depending on the product and workflow. Also record whether the system interacts directly with people, generates or manipulates content, recognises emotions or biometric categories, or publishes material that could be understood as authentic.

• What role does the organisation have in this specific use case?
• Who sees or receives the output, and in what context?
• Is text, audio, image or video created or altered wholly or partly by AI?
• Who decides on labelling, approval and later correction?

Separate provider duties from visible notices

Providers of systems that interact directly with people generally need to inform them that AI is involved, unless this is already obvious to a reasonably informed and attentive person. Providers of generative systems must also mark certain outputs in a machine-readable format and make their artificial generation or manipulation detectable, as far as technically feasible.

A machine-readable mark supports technical detection. It does not automatically replace a visible disclosure that a deployer may need for a deepfake or certain public-interest text. Document separately which signal the system produces and which notice people actually see in each channel.

Plan disclosure for deepfakes and public-interest text

Deployers must disclose artificially generated or manipulated image, audio or video content when it constitutes a deepfake. The European Commission guidelines say the notice should be clear, understandable and perceptible when people first encounter the content. Artistic, creative, satirical, fictional or similar works have specific rules that allow appropriate disclosure without unnecessarily hampering display or enjoyment.

Disclosure is also envisaged for AI-generated or manipulated text published to inform the public on matters of public interest. The statutory exception depends, among other things, on human review or editorial control and an identifiable person or legal entity carrying editorial responsibility. Simple proofreading should not be assumed to amount to substantive editorial control.

Account for the transition period and voluntary code

Article 50 duties generally apply from 2 August 2026. According to current EU information, a limited transition period until 2 December 2026 only concerns the marking and detection duty in Article 50(2) for certain systems placed on the market before 2 August 2026. It does not postpone every transparency duty.

The European Code of Practice on transparency of AI-generated content is voluntary. The legal duties remain binding. Eligible organisations may use the code as a structured route for demonstrating compliance; organisations choosing another route need to be able to explain its adequacy to the competent authority.

Build a reviewable workflow

Start with an inventory of relevant systems, roles, channels and content types. For each use case, decide who creates the label, who checks it before publication, how a technical signal is preserved and how missing or incorrect notices are corrected. Test the result separately on websites, apps, documents, video and social channels because a notice can be lost during export or upload.

Keep only the evidence required for your process and applicable retention rules. Useful records may include a template version, the approval step, the content type and evidence of testing. The responsible legal or specialist function must determine the exact legal assessment, documentation requirement and retention period.

Train roles using real publication workflows

Effective training should connect product, editorial, marketing, legal and operations roles. Use a realistic workflow: generate an AI output, determine the role, inspect the technical mark, draft the visible disclosure, record approval and verify the result in the target channel. Observable learning outcomes are more useful than promising automatic compliance after a course.

NextNowa can plan role-based AI governance training around your systems, channels and example workflows. Pricing and dates are agreed on enquiry. Before booking, the content, responsible instructors, availability and delivery scope are confirmed in writing.