Article 50 transparency obligations explained

Article 50 transparency obligations require four targeted disclosures or markings under the EU AI Act. Providers must inform users they are interacting with an AI system (unless obvious) and apply machine-readable marking to generated or manipulated content. Deployers must notify people exposed to emotion recognition or biometric categorisation systems and disclose the artificial origin of deepfakes or certain AI-generated text on matters of public interest. All information must be clear, accessible, and provided in plain language. These rules apply from 2 August 2026.[1][2]

They are designed to reduce deception and misinformation risks without imposing unnecessary burdens where context already makes the AI nature obvious. This page maps the obligations operationally, clarifies the provider-deployer split, and provides practical implementation steps, examples, and checklists.

Law status (May 2026) Article 50 of Regulation (EU) 2024/1689 is fixed current law and applies from 2 August 2026. A voluntary Code of Practice on marking and labelling of AI-generated content (covering key aspects of paragraphs 2 and 4) is in its second draft following stakeholder input, with finalisation expected in June 2026. The Digital Omnibus proposal and related negotiations focus primarily on high-risk system timelines and do not alter the core Article 50 transparency obligations. This page reflects the Regulation text and official Commission materials. It is not legal advice or a guarantee of compliance. Check the AI Act Service Desk timeline and EUR-Lex for authoritative updates.

Article 50 is not a single blanket requirement. It creates four distinct buckets with different triggers, actors, and methods. The information provided must always be clear, accessible, and in plain language (Article 50(5)).[1]

1. Interaction with AI systems Providers of AI systems intended to interact with natural persons must inform users that they are communicating with an AI system rather than a human, unless this is obvious from the circumstances and context of use.

2. Machine-readable marking of AI-generated or manipulated content Providers must ensure that outputs (text, image, audio, video) generated or manipulated by their AI systems are marked in a machine-readable format. This marking should enable detection tools and be effective, interoperable, robust, and reliable as far as technically feasible, taking into account content type, costs, and state of the art. The voluntary Code of Practice provides further detail on metadata, watermarking, fingerprinting, and logging approaches.[2]

3. Notice for emotion recognition or biometric categorisation exposure Deployers of AI systems for emotion recognition or biometric categorisation must inform natural persons exposed to the system.

4. Deepfake and public-interest text disclosures Deployers must disclose when content is a deepfake (artificially generated or manipulated image, audio, or video that resembles real persons, objects, places, or events in a way that would appear authentic) or when AI-generated or manipulated text is published on matters of public interest. This does not apply where the content has undergone human review or is subject to editorial responsibility. The Code of Practice addresses artistic, creative, satirical, or fictional contexts with more flexible regimes.[2]

This table operationalises the requirements drawn from the Regulation and Commission FAQ.[1]

The AI Act maintains a clear provider versus deployer distinction that determines responsibilities under Article 50.

• Providers develop the AI system (or have it developed) and place it on the market or put it into service. They bear the interaction notice duty (paragraph 1) and the machine-readable marking duty for generative outputs (paragraph 2). For generative AI tools, this often means implementing technical marking solutions (metadata, watermarks) that downstream deployers can build upon.

• Deployers (often called users or operators) use the AI system under their own responsibility for a specific purpose. They owe the notice duties for emotion recognition/biometric systems (paragraph 3) and for deepfakes or public-interest AI text (paragraph 4).

Generative AI providers and deployers diverge sharply. A company that builds and offers an image-generation API or chatbot model is typically a provider and must focus on machine-readable marking. An organisation that integrates that model into its customer-support workflow or newsroom publishing pipeline is usually the deployer and must handle visible disclosures for deepfakes or public-interest articles (unless human editorial processes apply). Overlap can occur when a single organisation both develops and operates the system; in that case, it must meet both sets of duties.[2]

See also: Chatbots and the EU AI Act for practical role mapping in conversational tools and Marketing agencies, publishers, and AI-generated public content for content workflows.

Effective implementation means useful, noticeable information that people actually see and understand — not hidden footnotes or purely technical checkboxes.

Placement, wording, and accessibility Notices should appear where and when the user needs them. Use plain language. Support accessibility standards so notices are readable by screen readers and available in appropriate languages. The Code of Practice (second draft) emphasises design and placement requirements for icons, labels, and disclaimers to promote consistency while allowing context-specific solutions.[2]

Practical approaches

• Customer-support chatbot example: Place the notice in the chat header and repeat it if the conversation shifts to sensitive topics. Keep a version history of the exact wording used.
• Image generation tool example: Combine visible on-image labelling with machine-readable metadata. Provide users a way to download a "transparency package" with provenance information. This satisfies both provider marking duties and helps deployers meet any further obligations.
• Newsroom or publisher using AI-generated public-interest text: Implement a workflow checklist that flags AI-assisted sections. If human editors review, revise, and take editorial responsibility, the disclosure exception can apply. Otherwise, add a clear statement such as "Parts of this article were generated by AI and have not received full human editorial oversight."

Metadata, watermarking, or cryptographic provenance should follow emerging standards referenced in the Code of Practice. Maintain technical documentation showing how your chosen methods meet the "effective, interoperable, robust and reliable" criteria as far as technically feasible. Test detection with available tools and keep logs. Avoid "theatre" — a watermark that no one can detect or a notice no one sees does not meet the operational goal of informed users.[2]

The law builds in targeted exceptions to avoid pointless disclosures.

• Obvious interaction: No notice is required if the AI nature is clear from the circumstances and context (for example, a clearly labelled industrial robot arm or a synthetic voice explicitly marketed as artificial in a creative audio project).
• Human review and editorial responsibility: Deployers of AI systems producing text publications on matters of public interest are exempt from disclosure if the content undergoes human review or falls under established editorial responsibility. The second draft Code of Practice clarifies reliance on existing journalistic or publishing processes.
• Creative, satirical, artistic, or fictional contexts: The draft Code recognises specific regimes for these uses, reducing burdens where audiences are not likely to be deceived about authenticity. Always document your reasoning for relying on an exception.[2]

Edge cases require judgment. A customer-support chatbot that occasionally generates creative responses is still primarily an interaction system. An image generator used internally for mock-ups may have lighter obligations than one publishing deepfakes for marketing. When in doubt, err toward transparency while documenting your assessment. The Commission’s Article 50 text helper and forthcoming guidelines provide additional clarification.

• Treating Article 50 as one generic "AI disclosure" rule instead of mapping uses to the four specific buckets.
• Burying notices in terms of service or footers instead of placing them where users actually interact with the output.
• Implementing only invisible technical marking without visible labelling where deployer duties apply.
• Failing to distinguish provider marking obligations from deployer disclosure obligations in generative AI supply chains.
• Over-disclosing in obvious contexts or under-disclosing public-interest content without qualifying for the human-review exception.
• Ignoring accessibility — notices that cannot be read by assistive technology or non-native speakers.
• Not maintaining evidence (screenshots, technical specs, workflow logs) that demonstrates how obligations are met.

• Inventory all AI systems and map each use case to the four buckets using the obligation map above.
• Assign clear owner roles for provider-style technical marking versus deployer-style visible notices.
• Design notices and labels that are prominent, plain-language, accessible, and context-appropriate; test with real users.
• For generative systems, implement and document machine-readable marking (metadata/watermarking) aligned with the latest Code of Practice draft.
• Create version-controlled templates for common disclosures and integrate them into product workflows and publishing pipelines.
• For public-interest or deepfake content, establish human review processes where possible to qualify for exceptions and document decisions.
• Maintain evidence artefacts (UI screenshots, technical documentation, detection test results) for each system.
• Subscribe to official updates and review your approach when the final Code of Practice is published.
• Use a dedicated generator tool to produce consistent, versioned disclosure language tailored to your interfaces.

Ready to operationalise Article 50 transparency? Generate compliant, context-specific notices and labelling packages with the Article 50 Disclosure Generator.

See a concrete worked example in the chatbot transparency sample report.

Stay current with the latest draft developments at Article 50 transparency code second draft: what matters now.

Sources (official only)

• Regulation (EU) 2024/1689, Article 50 and Article 113 (EUR-Lex).
• AI Act Service Desk: Article 50 text helper and implementation timeline.
• European Commission Transparency FAQ and Code of Practice landing page (digital-strategy.ec.europa.eu).
• Second draft Code of Practice announcement (March 2026).