Which four obligations does Art. 50 AI Act contain?
Art. 50 contains four transparency obligations, two for providers and two for deployers. A provider is anyone who develops an AI system or has one developed and places it on the market or puts it into service under its own name; a deployer is anyone who uses an AI system under its own authority in a professional capacity (Art. 3(3) and (4)). For providers, the Act speaks of machine-readable marking, and for deployers of disclosure. In everyday language, both are called labelling.
| Obligation | Who | What | Typical case |
|---|---|---|---|
| Para. 1: notice in direct interaction | Provider | Anyone who talks or writes directly with the system learns that it is an AI, unless this is obvious. | Chatbot for customer queries on the website |
| Para. 2: machine-readable marking | Providers of generative systems | Generated text, images, audio and video carry a machine-readable marking, for example in metadata or as a watermark. Assistive standard editing, such as a spell check, is exempt. | An obligation of the makers of writing and image tools, not of their users |
| Para. 3: emotion recognition, biometric categorisation | Deployers | The persons concerned are informed about the operation of the system. | Software that infers emotions from callers' voices; emotion recognition of employees in the workplace is prohibited, except for medical or safety reasons (Art. 5(1)(f)) |
| Para. 4: deepfakes and certain texts | Deployers | Disclose that deepfakes have been artificially generated or manipulated and that published texts on matters of public interest come from an AI. | Promotional video with an AI person, blog post on a change in the law |
Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply […] where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.
Since when has the labelling obligation applied, and are there transitional periods?
Art. 50 has applied since 2 August 2026 (Art. 113 AI Act) to all systems concerned, regardless of when they came onto the market. The Digital Omnibus, Regulation (EU) 2026/1744 of 8 July 2026, inserted a single transitional period: providers of generative systems that were already on the market before 2 August 2026 have until 2 December 2026 for machine-readable marking. There is no grace period for deployers or for the chatbot notice.
Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026.
According to the Commission's guidelines, you do not have to label old content retrospectively if it was generated and published before 2 August 2026. Texts kept in reserve are a different matter: an AI text that was created earlier but only appears on or after 2 August 2026 falls under the obligation. For older, unlabelled deepfakes that you continue to distribute, the Commission recommends labelling, but does not expect disproportionate effort such as searching through old archives or changing packaging that has already been printed.
Who is obliged: your company, your agency or the vendor?
The obligations fall on the company, not on the individual at the desk. According to the guidelines, employees working under instructions are not deployers in their own right; the company also remains the deployer if freelancers or contractors operate the system on its behalf and under its control.
- You commission an agency: according to the guidelines, if you only commission and do not decide whether and how the agency uses AI, you are not a deployer. Still, set out in the contract who does the labelling.
- You use a ready-made tool: the chatbot notice and the machine-readable marking must be built in by its maker, so by OpenAI for ChatGPT and by Mistral AI for Mistral Vibe (overview: European ChatGPT alternatives). Your obligations under paras. 3 and 4 remain.
- You build it yourself: if you develop a chatbot in-house or modify someone else's system, for example with your own training data, and put it into service under your own name, you are a provider and must ensure the notice under para. 1 yourself.
- Private use: if you use AI purely personally and not professionally, you have no deployer obligations (Art. 2(10)). If you regularly earn money with it, you are acting professionally.
What do SMEs have to label in typical cases?
For most office tasks, the answer is: nothing. You have to label if you use deepfakes, that is, realistic images, videos or voices that resemble real people, objects, places or events, or if you publish unreviewed AI texts on matters of public interest. The table classifies common cases according to the Commission's guidelines of 20 July 2026; in the individual case, content, audience and context matter.
| Case | Label? | Reason |
|---|---|---|
| Website chatbot answers customer queries | Yes, notice at the start of the conversation | Customers may take the answers for human ones (para. 1). With a purchased product, the vendor builds in the notice; check that it appears. |
| Internal AI assistant for trained employees | No | The guidelines cite it as an example where the AI is obvious. |
| Email or quote drafted with AI, checked and sent by a person | No | Correspondence with individual recipients does not count as published, and the customer is writing to a person, not to the AI. |
| AI agent sends emails to customers on its own | Yes, in the agent | It interacts directly with people and should identify itself as an AI and say on whose behalf it acts (para. 1). |
| Letter with a professional recommendation to a single customer | No | Advice to a single client is not a published text; the guidelines give the example of a consultant. |
| Product description or advertising copy | Generally no | Does not fall under para. 4 as long as it contains no statements on, for example, health, consumer safety or sustainability. |
| Blog post or newsletter about a change in the law, health or the environment | Generally yes, unless reviewed for content with named responsibility | Published information on a matter of public interest (para. 4, second subparagraph). A newsletter is published if an indefinite, larger group can receive it, including by subscription. |
| AI product photo that shows the goods as more attractive or different from reality | Yes | A deepfake example from the guidelines; there is also a risk of misleading advertising under § 5 UWG (German Act against Unfair Competition). |
| Real product photo against an AI background, colour correction, cut-out | Generally no | According to the guidelines, such edits hardly influence perception, as long as nothing is misleading about the goods. |
| Photorealistic AI person or synthetic influencer in advertising | Generally yes | The guidelines list a realistic synthetic influencer testing a real product as a deepfake. |
| Empty flat furnished with AI in photos | Yes | The Commission uses this case, in connection with the EU icons, as an example of content partly altered with AI. |
| Video with an AI avatar of the managing director for the workforce | Yes, before playback | A deepfake example from the guidelines; for purely internal use, the Code of Practice considers a clear notice before viewing sufficient. |
| Illustration, comic, recognisably fantastical motif | No | Anything that cannot appear real, such as quarrelling mice in a cheese advert, is not a deepfake according to the guidelines. |
When does the labelling requirement for AI texts not apply?
The obligation for AI texts does not apply if two conditions are met together: a person has reviewed the content, and someone holds editorial responsibility. For companies publishing their own specialist articles, this is often easier than a label. The guidelines set three requirements:
- Review the content: one or more people with relevant expertise engage with the content; fact-checking is the minimum. A spell check, approval without engaging with the content, or a review by a second AI is not enough.
- No more AI after approval: if you rephrase or add to the content of the text with AI after approval, you lose the exemption.
- Make responsibility visible: the name or role and contact details of the responsible party should be published where they are easy to find, online for example in the terms of use or other legal information such as the legal notice (Impressum).
- Assign responsibility
Decide who approves the content of specialist articles, newsletters and press releases, and state this role with contact details somewhere easy to find, such as the legal notice.
- Check the facts
Verify figures, dates, quotations and references against the source. Whatever cannot be substantiated comes out.
- Record the approval
Note the date, the person and the approved version. According to the guidelines, anyone who does not sign the Code of Practice should be able to show in another way how they meet the obligation.
What does adequate labelling look like?
The notice must be clear and distinguishable, appear at the latest at the time of first interaction or exposure, and meet accessibility requirements (Art. 50(5)). People must be able to perceive it without technical aids; a machine-readable marking by the vendor is therefore not enough for deployers. The guidelines name as inadequate:
- a notice only in the terms and conditions, terms of use, manual or behind several menu levels,
- a general sentence such as “Services on this website use AI” or the mere label “assistant”,
- a technical statement such as “this system uses a language model” that does not say that an AI is answering.
For deepfakes and published texts, the EU provides freely usable icons: a basic icon with the abbreviation “AI” and versions for fully AI-generated content and for content partly altered with AI. The icons are voluntary, the labelling is not, and an icon alone does not prove that the obligation is met. The Code of Practice suggests placing it, for example, in the top right-hand corner of an image or video, and for texts in the area above the text or near the headline.
What does a breach of Art. 50 cost, and who checks?
- €15 million · 3%
- Maximum fine for breaches of Art. 50; the higher amount applies (Art. 99(4)(g))
- SMEs
- pay at most the lower of the two amounts (Art. 99(6)), and since the Omnibus so do small mid-cap enterprises (para. 6a)
- €60,000
- Maximum for an SME with €2 million annual turnover: 3% of it, because the lower amount applies
The percentage refers to worldwide annual turnover in the preceding financial year. In Germany, the Federal Network Agency (Bundesnetzagentur) has been the competent market surveillance authority since 29 July 2026, unless the KI-MIG (German AI Market Surveillance and Innovation Promotion Act) provides otherwise (§ 2). The market surveillance authorities also impose the fines under Art. 99; the procedure follows the Act on Regulatory Offences (§§ 16 and 17 KI-MIG). Anyone can lodge a complaint, centrally with the Federal Network Agency (§ 8 KI-MIG, Art. 85 AI Act). If an AI system is based on a general-purpose model from the same provider, such as a chatbot built on the provider's own language model, the Commission's AI Office supervises the provider obligations instead, according to the Commission.
According to the guidelines, anyone who does not sign the Code of Practice should show in another way how they meet the obligations, for example by comparing their own measures with those of the Code, and may receive more requests for information. Authorities may take compliance with a Code assessed as adequate into account as a mitigating factor when setting the amount of a fine.
Guidelines and Code of Practice: do you have to sign anything?
No. On 20 July 2026, the Commission published guidelines on Art. 50, 51 pages with many examples. They are not binding; only the Court of Justice of the EU can interpret the Regulation authoritatively. They will be formally adopted, and according to the Commission only then become applicable, once all language versions are available; on 23 September 2026, only the English version existed. The Federal Network Agency offers a summary in German.
Since 10 June 2026, there has also been the Code of Practice on Transparency of AI-generated Content, with one part for providers (marking and detection) and one for deployers (labelling). The Commission assessed it as adequate on 8 July 2026, and the AI Board did so the following day. Signing is voluntary; by the end of July 2026, according to the Commission, around 190 companies and organisations had done so. According to the Commission, the Code does not provide conclusive proof that the obligations are met, even for signatories.
Checklist: putting AI labelling into practice in your business
- Take stock: where does your business generate or alter images, videos, voices or published texts with AI, and where do customers talk directly to an AI?
- Check the chatbot: is the AI notice visible in the chat window at first contact, and not just in the terms and conditions? If you built or modified it yourself and run it under your own name, you are the provider.
- Classify images and videos: does something look real and show people, products, places or events differently from how they are? Then label it and check separately for misleading content under the UWG.
- Define the editorial process: who reviews the content of specialist articles and newsletters, and which role is named for this in the legal notice?
- Decide on the label: EU icon or your own clearly legible notice, in a fixed place in the image or above the text.
- Contracts and training: agree with agencies who does the labelling; include labelling in the AI literacy training under Art. 4.
- Set a review date: the Commission intends to revise the guidelines when practice and case law require it. Check your classification again then.
The other dates of the EU AI Act are in the overview EU AI Act: obligations and deadlines, the training obligation in the article AI literacy under Art. 4, and what makes an AI agent in AI agents in the workplace.
Frequently asked questions
Do I have to label texts on my website that were written with AI?
Only if the text informs the public on a matter of public interest, such as politics, health, the environment or consumer safety, and nobody has reviewed its content. If a person with relevant expertise reviews the content and someone holds editorial responsibility, the obligation does not apply. Product descriptions and advertising copy generally do not fall under it.
Do I have to label AI-generated images?
Yes, if the image is a deepfake: realistic, resembling a person, object, place or event, and liable to be falsely taken as genuine. You do not have to label recognisably artificial illustrations, and colour corrections or a new background behind a real product are generally not a deepfake, as long as nothing is misleading about the goods.
Does the chatbot on our website need an AI notice?
Yes, if customers could take the answers for human ones; the guidelines expressly mention customer service chatbots. The obligation falls on the provider of the system, so if you built the chatbot yourself, on you. The notice belongs visibly at the start of the conversation; a sentence in the terms and conditions is not enough.
Do emails and quotes written with AI have to be labelled?
No. Business correspondence with individual recipients does not count as published, and if a person checks and sends the draft, the customer is not writing directly to an AI. It is different if an AI agent writes and sends emails on its own: it must identify itself as an AI.
Is a notice in the legal notice or the terms and conditions enough?
Not as labelling. The notice must be clearly recognisable at the latest at the first contact with the content or the chatbot; the guidelines expressly name notices only in terms of use or manuals as inadequate. Legal information such as the legal notice (Impressum) is, however, the right place to state who holds editorial responsibility for reviewed texts; the guidelines require this for the exemption for reviewed texts.
Does old AI content have to be labelled retrospectively?
No, anything generated and published before 2 August 2026 does not have to be labelled retrospectively. A text created earlier that only appears afterwards does, however, fall under the obligation. For older deepfakes, the Commission recommends labelling, without expecting disproportionate effort.
Sources
- Regulation (EU) 2024/1689 (AI Act), Art. 2, 3, 5, 50, 85, 99 and 113
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) of 8 July 2026, Official Journal of 24 July 2026
- European Commission: Guidelines on transparency obligations for providers and deployers of AI systems, 20 July 2026
- European Commission: Transparency obligations under Article 50 of the AI Act (questions and answers, as of 24 July 2026)
- European Commission: Code of Practice on Transparency of AI-generated Content
- European Commission: Opinion on the assessment of the Code of Practice, 9 July 2026
- European Commission: EU Icons for labelling AI-generated content
- Federal Network Agency: transparency obligations under Art. 50 AI Act
- Federal Network Agency: summary of the guidelines on transparency obligations
- Federal Network Agency: practical guide to the transparency of AI-generated content
- German AI Market Surveillance and Innovation Promotion Act (KI-MIG) of 22 July 2026, §§ 2, 8, 16 and 17
- § 5 UWG: misleading commercial practices
This text is not legal advice. It is the groundwork we had to do for ourselves, with the legal references, so that your lawyer does not have to start from scratch. Where a question depends on your circumstances, the text says so.