Article 50 code second draft: what matters now

The second draft of the Article 50 Code of Practice (published 5 March 2026) simplifies marking, labelling and design requirements for AI-generated content. It introduces greater flexibility for providers and deployers while keeping the core transparency obligations of Article 50 unchanged. These obligations become applicable EU law on 2 August 2026.

Current law requires providers of certain generative AI systems to mark outputs in a machine-readable way so they can be detected as artificial. Deployers must label deepfakes and certain AI-generated text on matters of public interest (with defined exceptions). The Code of Practice remains voluntary: it is a practical tool to help organisations demonstrate they meet those legal duties. The March 2026 draft narrows technical complexity, promotes open standards and an EU icon, and removes earlier taxonomy that distinguished “AI-generated” from “AI-assisted” content. Teams that begin preparing metadata workflows, label placement rules, archiving processes and editorial guidance now will be better positioned when the final code appears in June 2026 and obligations apply in August.

Law-status box Current law (applies 2 August 2026): Article 50 transparency obligations for providers and deployers of specific AI systems (interactive, generative, deepfake and public-interest text). See official timeline. This document: Second draft of a voluntary Code of Practice. Feedback closed 30 March 2026; final version expected beginning of June 2026. The draft does not change the law. It offers one route to show compliance. No certification is available or promised.

The European Commission published the second draft of the Code of Practice on Marking and Labelling of AI-generated content on 5 March 2026. The draft incorporates stakeholder feedback gathered through an EU survey, January 2026 workshops, Member State input via the AI Board, and European Parliament observations.

• Feedback window: Open until 30 March 2026 (now closed).
• Expected finalisation: Beginning of June 2026.
• Application of underlying legal rules: 2 August 2026 (Article 50 transparency obligations).

The code is structured in two sections that mirror the legal obligations:

• Section 1 (Providers) covers obligations under Article 50(2) for generative AI systems. It focuses on technical marking and detection of outputs (audio, image, video, text) so they are identifiable as artificially generated or manipulated.
• Section 2 (Deployers) addresses Article 50(4) obligations for labelling deepfakes and AI-generated or manipulated text publications on matters of public interest (unless the text has undergone human review and editorial responsibility).

Cross-cutting issues include clear, accessible information to users (Article 50(5)) and cooperation along the value chain. The Commission is preparing separate guidelines on the full scope of Article 50 in parallel. The code itself is voluntary; organisations that follow an approved final version can use it as evidence of good-faith compliance efforts.

The update page you are reading translates the policy announcement into concrete provider and deployer actions you can take before the final code and the 2 August 2026 deadline. For the underlying legal text, see the official Article 50 helper.

The second draft is explicitly described by the Commission as “streamlined and simplified,” offering more flexibility and reducing compliance burden while maintaining technical feasibility. Key practical shifts include:

• Two-layered marking approach (Section 1): Secured metadata combined with watermarking is now the baseline. Optional fingerprinting and logging have been retained as complementary tools rather than mandatory layers. Emphasis is placed on open standards, interoperability, robustness and proportionality (taking cost, content type and state-of-the-art into account).
• Removal of taxonomy complexity (Section 2): The earlier distinction between purely “AI-generated” and “AI-assisted” content has been eliminated. This removes a major source of classification uncertainty for deployers.
• Task-force and EU icon ideas: Section 2 proposes a task force to develop a uniform, interactive EU icon. The annex already includes illustrative examples that signatories may use. Placement and design requirements for labels, icons or disclaimers are provided at a minimum level of uniformity while allowing adaptation to context.
• Greater clarity on exceptions: Regimes for artistic, creative, satirical, fictional works and human-reviewed/editorially-controlled text publications are better defined, enabling reliance on existing editorial workflows.
• Overall tone: Stronger focus on practicality, reduced prescription, and cooperation between providers and deployers across the value chain.

These changes respond directly to stakeholder feedback that the first draft was overly complex in places. The result is a narrower, clearer discussion of marking, labelling and design requirements ahead of the fixed 2 August 2026 transparency date in current law.

You do not need to wait for the final code. Many preparatory steps align with both the draft and the underlying Article 50 obligations that apply from August 2026. Focus on workflows you can implement or pilot today:

• Prepare notices and disclosure language: Draft clear, accessible statements that outputs are AI-generated. Test them in user interfaces, content footnotes and metadata. Link to the Article 50 Disclosure Generator for structured templates.
• Metadata and watermarking plans: Audit current generative pipelines for the ability to embed secured metadata and apply robust watermarks. Prioritise open standards. Begin testing detection mechanisms on your own outputs.
• Label placement rules: Define where and how labels, icons or disclaimers will appear for deepfakes and public-interest text. The draft’s design requirements (visibility, non-intrusiveness, uniformity) can guide internal style guides now.
• Archive workflows: Establish processes to retain copies or logs of marked content for accountability, especially for deepfakes or public-interest publications.
• Public-interest text editorial guidance: Create or update editorial policies that document meaningful human review. This helps organisations rely on the human-review exception in Section 2.
• Cross-functional alignment: Bring product, legal, design and content teams together to map provider/deployer responsibilities. Use our Article 50 transparency obligations explained for role-split checklists.
• Sector-specific preparation: Marketing and publishing teams should review how labelling affects campaign assets (/sectors/marketing-and-publishers). Chatbot operators should ensure users are informed they are interacting with AI (/sectors/chatbots).

These steps build operational readiness and create evidence of proactive compliance regardless of the exact final wording of the code.

Provider planning content marking. A mid-sized image-generation company reviews its pipeline against the two-layered approach in the second draft. It adds secured metadata fields and tests watermark robustness on JPEG and PNG outputs. It documents the technical choices and state-of-the-art considerations so it can update quickly once the final code is published. The company also joins the open-standards discussion to reduce future integration costs.

Publisher planning disclosure workflow. An online news platform that occasionally uses AI to draft summaries of public-interest reports updates its editorial handbook. Articles that receive substantive journalist review are flagged as “human-reviewed” and exempted from labelling. For fully AI-generated explanatory pieces, a standardised footer disclaimer is added automatically. The workflow logs the classification decision for audit purposes, directly supporting Section 2 of the draft.

Is the code mandatory? No. The Code of Practice is voluntary. Following an approved final version can help demonstrate compliance with Article 50(2) and (4), but it is not a legal requirement. The underlying transparency obligations in Article 50 are mandatory from 2 August 2026.

Should teams wait for the final code? No. The core legal obligations are fixed. The second draft already shows the direction of travel (flexibility, open standards, EU icon, simplified taxonomy). Starting metadata, labelling and editorial workflows now reduces risk and creates evidence of good-faith preparation. Waiting until June 2026 leaves only weeks before the August deadline.

• Treating the voluntary code as if it delays or replaces the 2 August 2026 legal deadline for Article 50.
• Over-engineering classification systems for “AI-generated” versus “AI-assisted” content after the second draft removed that taxonomy.
• Implementing rigid labelling that ignores the draft’s emphasis on flexibility and user experience.
• Failing to document human-review processes for public-interest text, thereby losing access to the editorial exception.
• Assuming technical marking is solely a provider responsibility—many deployers must still integrate, verify and present labels appropriately.
• Not piloting open standards or metadata solutions early, which may increase costs once interoperability expectations crystallise.

Action checklist

• [ ] Map your products and content types to Article 50 roles (provider vs deployer).
• [ ] Pilot metadata + watermarking on at least one content type.
• [ ] Draft and user-test labelling language and placement for deepfakes and public-interest text.
• [ ] Update editorial guidelines to capture meaningful human review where relied upon.
• [ ] Create an internal archive and logging policy aligned with the draft’s detection and verification expectations.
• [ ] Review outputs using the Article 50 Disclosure Generator.
• [ ] Subscribe to updates so you receive the final code and guideline announcements as soon as they appear.

Stay ready. Subscribe for alerts on the final code, supporting guidelines and enforcement developments. Test your current materials with the Article 50 Disclosure Generator or explore readiness by sector: Chatbots and the EU AI Act and Marketing agencies, publishers, and AI-generated public content. For the full legal foundation, see our guide to Article 50 transparency obligations explained.

Published 20 April 2026. This page reflects official Commission announcements and the second draft text available on digital-strategy.ec.europa.eu. It is not legal advice.