Why AI-generated content must be labelled
Since 2 August 2026, Article 50 of the EU AI Act has applied. Transparency about AI-generated content has therefore gone from good practice to a legal requirement. Providers of generative AI systems must mark their output so that it can be detected as artificially generated. Anyone who uses AI to create deepfakes or to publish AI-generated text on matters of public interest must now disclose this to the people exposed to the content.
Your organisation’s role determines which requirements apply to you. A provider is the party that develops or sells an AI system, for example a company that builds a chatbot or an image generator. A deployer is the party that buys in and uses the AI system in its own operations, for example for marketing or customer communication. Most companies are deployers rather than providers, which the section “Which requirements apply to you” further down covers in detail.
To make the requirements manageable in practice, the EU AI Office has developed a voluntary code of practice, the ‘Code of Practice on Transparency of AI-Generated Content’, which gives concrete guidance on how to comply with Article 50. There are also the Commission’s ‘Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act’, which are the Commission’s non-binding interpretation of Article 50. In this article, we help you sort out the concepts.
Background: the legal sources behind Article 50
The four legal sources
This article is based on four complementary sources. The foundation is Article 50 of the EU AI Act, which has been binding law since 2 August 2026. However, the article was amended by the Digital Omnibus package, which entered into force on 27 July 2026. In addition, there is the code of practice, which is voluntary but gives concrete guidance on marking and detection, and the Commission’s guidelines, which are a non-binding interpretation of the whole of Article 50.
Article 50 has three purposes: to protect people’s autonomy, to counter manipulation and misleading information, and to build technical trust in AI-generated content. By informing us that we are interacting with AI or are exposed to AI-generated content, it should help us make well-informed decisions and preserve our trust in the information ecosystem, while protecting freedom of expression and the credibility of the media.
The history and legal status of the code of practice
The code of practice was developed by two working groups under the EU AI Office: one for technical marking and detection solutions, and one for labelling deepfakes and AI-generated text. The final version was adopted in June 2026.
The European Commission and the Board (the European Artificial Intelligence Board, made up of representatives of the Member States) have assessed the code as an “adequate voluntary tool” for demonstrating compliance with Article 50(2) and 50(4) to (5) of the EU AI Act. Signing up is voluntary, but by the end of July 2026 around 190 organisations had already signed.
According to the Commission’s guidelines on Article 50 (July 2026), signing up to the code gives providers and deployers a simple, predictable and legally certain way to demonstrate compliance with Article 50(2), 50(4) and 50(5). Supervisory authorities then focus their checks on whether you actually follow the code in practice. Those who do not sign up must instead demonstrate compliance by “other appropriate means” and should expect more detailed questions and requests for information from the market surveillance authorities (for Article 50 in Sweden, these are PTS and IMY).
The code is divided into two sections:
- Section 1 is aimed at providers of generative AI systems and sets out requirements for technical marking and detection of AI-generated content.
- Section 2 is aimed at deployers of AI systems and sets out requirements for labelling deepfakes and AI-generated text published on matters of public interest.
The scope of Article 50
Article 50 of the EU AI Act consists of seven paragraphs in total:
Transparency in AI interaction and marking of AI-generated content for providers
- Information about AI interaction: Providers of AI systems intended to interact directly with people must design the systems clearly. The user must understand that they are talking to an AI, for example a customer service chatbot. Exceptions apply if this is already obvious to a reasonably well-informed person, and for systems authorised by law to be used for law enforcement.
The requirement also explicitly covers AI agents that interact directly with natural persons or act on someone else’s behalf, for example when making bookings or entering into contracts. They too must clearly disclose their artificial nature and on whose behalf they act. - Marking of AI-generated content: Providers of generative AI systems, including general-purpose AI systems, must mark their synthetic content in a machine-readable format, so that the content can be detected as artificially generated. However, the requirement does not apply to purely assistive editing or to law enforcement authorised by law.
Emotion recognition and deepfakes for deployers
- Emotion recognition and biometric categorisation: Anyone who uses emotion recognition or biometric categorisation systems must inform the people exposed to them. Personal data must also be processed in line with the GDPR and relevant data protection law. Exceptions apply to law enforcement authorised by law.
- Deepfakes and published text: Anyone who uses AI to create or manipulate image, audio or video content that constitutes a deepfake must disclose that the content is artificial. For artistic, satirical or similar works, however, a lighter, non-intrusive disclosure is enough.
The same applies to AI-generated text published to inform the public on matters of public interest. The exception is when a named person or organisation has held editorial responsibility and reviewed the text.
General requirements and codes of practice
- When and how the information must be given: All the information above must be given clearly and distinguishably. It must be provided at the latest at the first interaction or exposure and meet applicable accessibility requirements.
- Relationship to other rules in the EU AI Act: The requirements in Article 50 do not affect the requirements for high-risk AI systems covered by Annex III.
- Codes of practice and the Commission’s role. The AI Office was originally tasked with ensuring that codes of practice were drawn up, but since the omnibus entered into force on 27 July 2026, it is instead the Commission that has the role of assessing whether a code is adequate to ensure compliance. The reason for the change is that codes of practice have limited legal effect and do not give any presumption of compliance.
Which requirements apply to you as a provider or deployer?
Article 50 sets different requirements depending on whether you develop generative AI or use it in your operations.
| Role | Who is covered | Main requirements | Key exceptions |
|---|---|---|---|
| Provider of a generative AI system | Companies that build AI systems that generate synthetic audio, images, video or text | Mark output in a machine-readable format (digital signature + invisible watermark), provide a free detection solution, achieve interoperability by 2 February 2027 | Single-layer marking is enough for closed, physical products |
| Deployer of an AI system | Anyone who uses AI to create deepfakes or publish AI-generated text on matters of public interest | Label the content clearly, e.g. with the EU AI icon, at the latest at the first exposure | Art, satire and fiction require lighter labelling; text under human editorial control is exempt |
For most companies, it is enough to use the EU’s ready-made, freely available icon. It already meets the code’s requirements on design and placement, so you rarely need to develop your own solution.
What this means for your organisation
The vast majority of organisations are deployers rather than providers. Rather than building your own AI models, you use tools such as chatbots, image and video generators, or text generators for marketing, customer communication and reports.
In concrete terms, this means that an image or video that resembles real people, places or events and that is AI-generated or AI-manipulated counts as a deepfake. It must then be labelled. AI-written text published to inform the public on matters of public interest, for example news-like content or opinion pieces, must also be labelled. The exception is when a named person has held editorial responsibility and reviewed the text. If you also use a customer service chatbot, Article 50(1) applies as well: users must understand that they are talking to an AI.
If you build your own generative AI product, or build on an AI model and sell it on, you also become a provider. You are then also covered by the heavier technical requirements: marking, watermarking and your own detection solution.
How to label AI-generated content: five concrete steps
- Map where in your organisation AI generates or manipulates images, audio, video or text (e.g. marketing, customer communication, reports, social media).
- Identify which content counts as a deepfake, or as text on matters of public interest without human editorial control.
- Implement labelling: Use the EU AI icon or equivalent when publishing, following the code’s design and placement specifications.
- Update internal policies, appoint someone responsible for editorial control where needed, and train staff who work with AI-generated content.
- Document your assessment of compliance with the requirements, so that you can show it to market surveillance authorities on request.
Do you already have AI-generated content in production that was placed on the market before 2 August 2026? Then you have until 2 December 2026 to meet the marking requirement in Article 50(2). This is due to a transitional rule introduced by the Digital Omnibus package in summer 2026.
Also keep in mind the difference between the system and the individual piece of content: AI-generated content or deepfakes created before 2 August 2026 do not need to be labelled retroactively. But if you make that content public or distribute it on or after 2 August 2026, for example by republishing an older AI-generated image or article, it must be labelled at that point, according to the Commission’s guidelines on Article 50.
If you build a generative AI product yourselves, there is a sixth step: build in technical marking and a detection solution, and keep track of the interoperability deadline of 2 February 2027.
Penalties for non-compliance
Failure to comply with Article 50 can be fined up to EUR 15 million or up to 3 per cent of total worldwide annual turnover for the preceding financial year, whichever is higher. At the same time, market surveillance authorities assess the seriousness of the infringement case by case, and may count adherence to an appropriate code of practice as a mitigating factor.
Summary
The transparency requirements in Article 50 are not a future issue. They have applied since August 2026. The code of practice does offer a practical, EU-reviewed way to demonstrate compliance. But it requires you to actually map where AI is used in your organisation and build labelling into your workflows.
At AddComply, we help you map your AI systems, assess which transparency requirements apply to your organisation, and build a documented compliance process. We’d be happy to book a demo and show you our platform.
In addition to the EU AI Act, the following sources have been used:
- Code of Practice on Transparency of AI-Generated Content, European Commission
- Guidelines on transparency obligations for providers and deployers of AI systems (Article 50), European Commission
