Transparency obligation (AI Act Article 50)

What is the transparency obligation?

The transparency obligation is the part of the EU AI Act that says people have a right to know when AI is in the room. It lives in Article 50 and it has applied since 2 August 2026. Belgian press tends to call it the labelplicht, the labelling duty, and that name covers about half of what it does.

The rules are short, five paragraphs, and they do not depend on how risky your use case is. The high-risk obligations only bite if your application is on a specific list. Article 50 bites the moment an AI system talks to a person or produces text, images, audio or video. A webshop with a chatbot, a marketing team that generates visuals, a company blog written with a language model: all of them are in.

The duties are light. Say that it is AI, and make sure that generated content can be recognised as generated, by people where it matters and by software everywhere.

The five duties and who carries each one

Article 50 splits its duties between the provider, the company that builds the AI system and puts it on the market under its own name, and the deployer, the company that uses it under its own authority (personal, non-professional use does not count). An SME that puts a vendor's chatbot on its website is the deployer. The chatbot vendor is the provider. Which role you have decides which of the five duties are yours.

1. Tell people they are talking to an AI (provider). A system designed to interact with people has to be built so that the person knows they are dealing with AI, unless that is obvious to a reasonably well-informed and observant person. The provider builds the disclosure in.

2. Mark generated content in a machine-readable way (provider). A system that generates audio, image, video or text has to mark its output so that software can detect it as artificially generated or manipulated. This is the watermark and metadata layer, invisible to the reader, and it sits with the provider of the generation tool.

3. Label deepfakes (deployer). A deepfake, in the AI Act's definition, is AI-generated or manipulated image, audio or video that resembles existing people, objects, places, entities or events and would falsely appear authentic to someone looking at it. Whoever publishes such content has to disclose that it was generated or manipulated. That is a visible label, and it is the deployer's job.

4. Disclose AI-written text that informs the public (deployer). Text generated by AI and published to inform the public on matters of public interest has to be disclosed as AI-generated. The exemption for text that a person has reviewed is narrower than most people assume, see below.

5. Inform people subject to emotion recognition or biometric categorisation (deployer). If you run a system that reads emotions from someone's face or voice, or sorts people into categories from their biometric data, you have to tell the people it is applied to. Emotion recognition at work and in schools is banned outright except for medical or safety reasons, so this duty mostly concerns other settings.

In every case the information has to reach the person in a clear and distinguishable way, at the latest at the first interaction or exposure, and it has to be accessible to people with disabilities.

The exemptions, and how narrow they are

The article has a handful of exemptions, and the Commission's guidelines of 20 July 2026 read each of them strictly.

Obvious AI. The chatbot disclosure can be skipped when it is obvious that you are dealing with AI. The test is what an average, reasonably well-informed and observant person would notice, and the guidelines say to interpret it restrictively. A widget called Ask our AI assistant is obvious. A chat window with a photo and the name Sarah is not.

Human review or editorial control for text. AI-written public-interest text needs no label when a person has reviewed it and someone holds editorial responsibility for it. The guidelines want a deliberate examination of the substance by someone with relevant knowledge. A spell check or a skim does not count, checking the facts is described as the minimum, and the mere possibility that someone could have reviewed it does not count either.

Artistic, creative, satirical and fictional work. Deepfakes in an evidently artistic or satirical context still need disclosure, but in a way that does not spoil the work, for instance in the credits.

Assistive editing. The machine-readable marking duty does not apply to functions that do standard editing or that do not substantially change the input.

Machine-readable marks, the Code of Practice and the EU icon

The AI Act treats two layers as separate duties with separate owners.

The machine-readable mark is what the generation tool embeds so that detection software can flag content as synthetic. The Commission's Code of Practice on transparency of AI-generated content, published in its final form on 10 June 2026, describes the expected approach: digitally signed metadata plus an imperceptible watermark for image, audio and video, and a single layer for free-form text. Signing the Code is voluntary, but the Commission has said that following it is an adequate way to show compliance, and that non-signatories have to demonstrate compliance in another way. Close to 190 companies and organisations had signed by the end of July 2026.

The visible label is what a reader sees. The Commission published a set of EU icons built around the letters AI, with variants for content fully generated by AI and for existing content partially modified by AI, in all 24 official languages and free to use without attribution. Using the icons is optional. Labelling is not. The Code suggests a label in a top corner of an image or video, near the headline or in the colophon of a text, and a spoken line at the start of audio.

One date matters for the machine-readable side. The Digital Omnibus on AI, adopted in July 2026, gave generative AI systems that were already on the market before 2 August 2026 until 2 December 2026 to add the marking. Systems placed on the market after 2 August 2026 mark from day one. None of the other Article 50 duties got a delay. Content generated before the rules applied does not need a mark or label after the fact; what you publish from now on does.

Transparency obligation versus high-risk obligations

The two are often mixed up because both sit in the same regulation and both were scheduled for August 2026. The dimension that separates them is scope: how many companies they touch and how much they ask.

The high-risk obligations apply to a narrow list of use cases from Annex III, such as CV screening, credit scoring and worker monitoring, plus AI built into regulated products. If you are on the list, the duties are heavy: risk management, data governance, documentation, logging, human oversight, conformity assessment. The Digital Omnibus pushed the deadline for stand-alone Annex III systems to 2 December 2027.

The transparency obligation applies to nearly everyone who uses generative AI or a conversational system in a professional setting, regardless of sector or risk. The duties are light and they apply now. A company can be completely outside the high-risk regime and still owe a chatbot disclosure, a deepfake label and a note under an AI-written article.

Both can apply to the same system. A recruitment chatbot that screens candidates is high-risk and has to say that it is a bot. Meeting Article 50 does not exempt anyone from the AI literacy duty in Article 4 either; the guidelines say so explicitly.

Fines and who enforces them in Belgium

Breaching Article 50 falls under Article 99(4) of the AI Act: administrative fines of up to 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher. For SMEs, including start-ups, the rule flips and the lower of the two amounts is the ceiling. A company with one million euros of turnover therefore looks at a maximum of 30,000 euros, not 15 million.

Enforcement is national. Belgium has pointed to BIPT, the telecom regulator, as its market surveillance authority for the AI Act, with FOD Economie as the information point for companies, but at the time of writing the law that formalises this had not yet passed parliament. The duties apply regardless.

A checklist for a Belgian SME

Chatbot on your website or in your app. Add one line at the start of the conversation, in the language of the visitor: you are chatting with an AI assistant, and here is how to reach a person. Check the vendor's settings, because the provider builds the disclosure in but you decide whether it is visible. If the bot acts on behalf of your company, say so; the guidelines expect an agent to state who it acts for.

Sales or support voice bot. The same duty, spoken. The bot introduces itself as an automated assistant before the substantive part of the call. A human name and a natural voice without that line is exactly what the obvious-AI exemption does not cover.

Marketing images and video. A stylised illustration of a product is not a deepfake and needs no visible label; the machine-readable mark is the tool vendor's job. A photo-realistic image of a person, a location or an event that someone could take for real is a deepfake and needs a label from you. When in doubt, use the EU icon in a corner.

AI-written blog posts and news items. If the text informs the public on a matter of public interest, you either label it or you have a named person review the substance and take editorial responsibility. Product copy on your own site is usually outside that scope, a piece on new energy subsidies or a trend in your sector is not. Keep a record of who reviewed what.

Vendor questions. Ask every generative AI supplier whether its outputs carry a machine-readable mark, whether it has signed the Code of Practice, and if its system was on the market before 2 August 2026, when it meets the 2 December 2026 date. Put the answers in your AI register next to that tool.