What to do before 2 August 2026 for Article 50 transparency: a checklist

The transparency obligations under Article 50 of the EU AI Act take effect on 2 August 2026 and are not postponed; this checklist walks through what to put in place per obligation, from chatbot disclosure and machine-readable marking to vendor contract clauses and an evidence file.

The 2 August 2026 deadline for Article 50 is fixed. Unlike the high-risk regime, which the political agreement on the Digital Omnibus moves to 2 December 2027, the transparency obligations are not shifted. For the background to that deadline and its relationship with the Digital Omnibus, see Article 50 transparency obligations: the deadline of 2 August 2026 that is not postponed and the Digital Omnibus and the postponement of the high-risk obligations.

This post is structured differently. It is not an explanation of the why, but an operational checklist of the what and the how. Each obligation gets concrete actions, sample text and points of attention, so you can turn today's inventory into a demonstrable implementation before the deadline.

Obligation 1: chatbot and AI-interaction disclosure (Article 50(1))

An AI system intended to interact directly with people, such as a chatbot, voice assistant or automated phone line, must be designed so that the person knows they are dealing with an AI system. The notice is not required where it is already obvious to a reasonably observant person.

Concrete actions:

• Inventory every customer- or employee-facing channel where an AI system communicates directly: website chatbots, in-app assistants, WhatsApp bots, voicebots and automated email handling.
• Place the disclosure at the first interaction, in a clear and distinguishable spot, and make it accessible to people with disabilities.
• Show the notice visibly in the opening message of the chat, not buried in a terms-and-conditions or privacy page.
• For voicebots: make the notice audible at the start of the call.

Sample disclosure text (chat): "You are chatting with an AI-based virtual assistant. Want to speak to a person? Type 'agent'." For a voicebot, a spoken variant at the start of the call suffices.

Note: the exception for the obvious case is narrow. An avatar or a name like "AI assistant" does not automatically make it obvious. Document, per channel, why you do or do not show an explicit notice.

Obligation 2: marking deepfakes and synthetic content (Article 50(2))

Generative AI systems that produce audio, image, video or text must mark that output in a machine-readable format and make it detectable as artificially generated or manipulated. This is technically the most demanding obligation, because it touches watermarking and provenance standards. An exception applies to systems with a purely supportive, editorial function, such as a grammar corrector.

Concrete actions:

• Map which systems generate or edit content: image generators, video and voice tools, and text generation in production flows.
• Implement a machine-readable marking. C2PA Content Credentials is the most mature open standard for provenance metadata in image, audio and video; for text you can include metadata or a standardised marking in the output.
• Where possible, combine the machine-readable marking with a visible indication, so the information is available to both systems and people.
• Test whether the marking is robust against common edits such as compression, cropping and conversion, and record the result of that test.
• Follow the upcoming Code of Practice on marking and labelling of AI-generated content from the European Commission; it offers providers a practical route to demonstrably meet 50(2).

Obligation 3: information duty for emotion recognition and biometric categorisation (Article 50(3))

Anyone deploying a system for emotion recognition or biometric categorisation must inform the exposed persons. This obligation sits with the deployer.

Concrete actions:

• First check whether the application does not already fall under a prohibition in Article 5. Emotion recognition in the workplace and in educational institutions is in principle prohibited, with a narrow exception for medical or safety reasons. A prohibited application is not solved with an information duty.
• Inventory where you process biometric signals to infer emotions, mental states or group characteristics, even if the vendor calls it "engagement", "attention" or "well-being". The substantive function is decisive, not the marketing term.
• Inform the persons concerned in advance, clearly and in an accessible way, about the use of the system.
• Place the information duty alongside your GDPR obligations, because biometric data is special-category personal data. The AI Act sits on top, not instead.

Obligation 4: AI-generated public-interest text and deepfakes (Article 50(4))

This obligation has two branches and sits with the deployer. Anyone generating or manipulating a deepfake must disclose that the content is artificial; for artistic, creative, satirical or fictional work a lighter form applies that must not hamper the enjoyment of the work. In addition, AI-generated text published to inform the public on matters of public interest must be marked as such, unless the text has been reviewed under human editorial responsibility.

Concrete actions:

• Mark deepfakes and manipulated image or audio visibly as artificial, for example with a text label at or within the publication.
• Review your publication flows for texts on matters of public interest: news-like items, public information and public communication. Mark AI-generated texts, unless a person has substantively edited the text and takes responsibility for it.
• Record when the editorial exception applies, so you can demonstrate per publication that a person reviewed the piece.

Sample marking text: "This image was generated or edited with AI." For public text without human editing: "This text was generated with the help of AI."

Ongoing prerequisites

Beyond the four obligations there are three organisational prerequisites that cut across all obligations and that you must not skip.

Vendor and contract clauses

Anyone integrating an external model or external service into their own product often sits in a mixed provider-and-deployer position. So contractually record who delivers which transparency obligation.

• Ask vendors to demonstrate that their generative output contains a machine-readable marking under 50(2), and which standard they use.
• Include a clause obliging the vendor to maintain the marking and disclosure functionality, also after model updates.
• Record that the vendor informs you in good time about changes that affect transparency.
• Avoid relying on a loose claim such as "AI Act compliant"; ask for the substantiation per obligation.

Internal accountability

• Assign an owner per system who is responsible for the transparency obligation.
• Embed AI literacy among the people who work with these systems; Article 4 on AI literacy has applied since 2 February 2025 and forms the human foundation under transparency.
• Make transparency part of your procurement and release process, so new systems do not go live without disclosure.

Documentation and evidence file

Article 50 does not require a formal conformity assessment, but a supervisory authority will, on a complaint, want to see that the disclosure was there and how it was designed. A simple evidence file is the difference between demonstrable compliance and reconstructing after the fact.

• Keep screenshots of chatbot disclosures and visible markings.
• Document the chosen marking standard and the robustness test results.
• Record per system the provider-or-deployer role split, plus the reasoning behind any exceptions.
• Keep the contractual arrangements with vendors together in a single file.

The checklist at a glance

Anyone who has put these four obligations and the three prerequisites in place before 2 August 2026 has the most visible layer of the AI Act in order. The technical marking under 50(2) deserves the most attention, because watermarking and provenance signals are not in place overnight. Start there first.