Article 50 has moved from preparation to practice
From 2 August 2026, the EU AI Act's Article 50 transparency obligations apply to specified AI systems and uses. The European Commission published final implementation guidelines in July, alongside a Code of Practice intended to help organisations meet the marking and labelling requirements for AI-generated or manipulated content.
The practical point is simple: organisations now need to know when their AI systems must disclose that a person is interacting with AI, when generated outputs need machine-readable marking and when people must be told that content or analysis involves AI.
This is not only a product-interface task. Compliance depends on accurate system inventories, clear provider and deployer roles, release controls, supplier evidence and records showing why a particular disclosure approach was selected.
Which business activities should be checked first
Start with AI systems that interact directly with people. Providers of interactive AI systems generally need to design them so individuals are informed that they are interacting with AI, unless that fact is obvious to a reasonably informed and observant person in the circumstances. Customer-service assistants, candidate-facing recruitment tools and automated advisory interfaces deserve early attention.
Next, identify systems that generate or manipulate image, audio, video or text. Providers within scope need technical measures that mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated. The Commission says those measures should be effective, interoperable, robust and reliable as far as technically feasible.
Deployers also have duties in specific contexts. These include disclosure when using emotion recognition or biometric categorisation systems, labelling deepfake content, and identifying certain AI-generated or manipulated text published to inform the public on matters of public interest. The exact rule and exceptions depend on the use case, so record the rationale rather than relying on a blanket statement.
The practical Article 50 evidence checklist
- Inventory the relevant systems. Record chatbots, virtual assistants, synthetic-media tools, content generators, emotion-recognition systems and biometric-categorisation uses, including embedded supplier features.
- Decide and record your role. State whether the organisation is acting as provider, deployer, importer or distributor for each use. Obligations follow the facts, not the internal product label.
- Map each transparency trigger. Link the system to the applicable interaction, content, deepfake, public-interest text or other disclosure requirement, plus any relied-on exception.
- Design the disclosure. Keep the approved wording, placement, timing, language, accessibility treatment and interface screenshots. A buried generic disclaimer is not a controlled measure.
- Implement machine-readable marking. For relevant AI-generated or manipulated outputs, record the technical method, interoperability assumptions and what happens when content is exported, compressed or shared downstream.
- Test the measure. Verify chatbot notices before or at interaction; labels across devices and formats; deepfake disclosures where people encounter the content; and the detectability, robustness and reliability of technical markings.
- Put people around the control. Name who approves wording, reviews ambiguous cases, handles exceptions and complaints, and can stop publication or release when the evidence is incomplete.
- Retain a versioned evidence pack. Keep scope decisions, screenshots, test outputs, supplier confirmations, approvals, exceptions, review dates and change history together.
Do not treat the voluntary code as a substitute for the law
The Commission describes the Code of Practice as a voluntary framework that can help signatories demonstrate compliance with specified transparency obligations. It does not replace the AI Act or the Commission's Article 50 guidelines.
Organisations that do not follow the code remain responsible for compliance and should be able to explain how their chosen measures satisfy the regulation. That makes a documented gap assessment useful even where a business decides not to sign or follow every measure in the code.
For generative AI systems placed on the market before 2 August 2026, check the applicable transitional position carefully. Do not assume that every existing system or every output has the same treatment.
What good evidence looks like now
A credible Article 50 evidence pack connects each relevant system to its role assessment, transparency trigger, implemented measure, test record, owner and review date. It should also show how supplier changes and product updates trigger reassessment.
This evidence can serve several purposes at once: regulatory readiness, customer assurance, procurement responses and internal release approval. The goal is not a one-off label audit. It is a repeatable process that keeps transparency measures accurate as systems and uses change.
AI Act Ready helps organisations organise this work into a living inventory, role map, control set and buyer-ready evidence pack. For the documents customers are likely to request, read our procurement checklist next.
Frequently asked questions
Does Article 50 apply to UK businesses?
It can. A UK business may be in scope when it provides or deploys an AI system in the EU, supplies an EU customer, or its system output is used in the EU. The precise position depends on the role, system and use case.
What evidence should an organisation keep for Article 50?
Keep the role and scope decision, disclosure wording, screenshots, machine-readable marking method, test results, human-review records, supplier evidence, approvals, owners and version history.
Is a general AI disclaimer enough?
Usually not. The disclosure must match the relevant system and use, be clear and timely, and be supported by evidence showing that the implemented measure works across the channels and formats in which people encounter it.
Turn the checklist into controlled evidence
Use the AI Act readiness check to identify your first gaps, see how the AI Act Ready software keeps evidence linked to systems and owners, or talk to us about an Article 50 evidence review.
This article provides practical governance information, not legal advice. Confirm the application of Article 50 to your specific systems and markets.