Nederlands

AI Act transparency rules: what 2 December 2026 means for AI content

From 2 December, older generative AI systems must also mark their output. What Article 50 of the AI Act asks of chatbots, AI images and AI-written text.

Koen Dercksen4 min read

The transparency rules of the EU AI Act have applied since 2 August 2026, but one part had a grace period. Generative AI systems that were already on the market before that date get until 2 December 2026 to mark their output as AI-generated. That is in under two weeks. If your company runs a chatbot, generates images or writes text with a language model, now is a good moment to check what Article 50 asks of you, and who in the chain has to do it.

What changes on 2 December

The Digital Omnibus on AI, the amendment to the AI Act that was published in July, added a new paragraph to Article 111. Providers of AI systems “generating synthetic audio, image, video or text content” that were placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026. Newer systems had no grace period.

Four duties in Article 50

Article 50 of the AI Act splits the work between the provider, who develops an AI system or has one developed and puts it into service under its own name, and the deployer, who uses it professionally.

Who What Article
Provider Tell people they are talking to an AI system, unless that is obvious 50(1)
Provider Mark generated audio, images, video and text in a machine-readable way, so it can be detected as AI-generated 50(2)
Deployer Disclose that a deepfake image, video or audio is artificially generated or manipulated 50(4)
Deployer Disclose AI-generated text published to inform the public on matters of public interest 50(4)

The marking duty in 50(2) does not apply when the system only assists with standard editing or doesn’t substantially change what the user put in. The text duty in 50(4) does not apply when the text has gone through human review or editorial control and a person or company holds editorial responsibility for the publication. Under Article 50(5), the information must be clear and given at the latest at the first interaction or exposure.

Are you a provider or a deployer?

That depends on what you built. If you use ChatGPT or an image tool as it comes, you are a deployer, and the model vendor is the provider. If you built a customer chatbot or a content tool on top of a model API and offer it under your own name, the definition in Article 3(3) can make you the provider of that system. Then the chatbot disclosure in 50(1) is your job, and so is the marking in 50(2) if the system generates content.

The line isn’t always clear. The Commission’s guidelines on Article 50 explain the roles along the value chain.

The code of practice and the EU icons

The Commission published a code of practice on 10 June 2026. It has a section for providers (marking and detection) and one for deployers (labelling deepfakes and AI-generated text). Signing is voluntary, and about 190 organisations had signed by the end of July. Companies that don’t follow the code must show in another, equally adequate way that they mark and label correctly. The obligations themselves are law.

For labels, the Commission made a set of EU icons: a basic icon, one for fully AI-generated content and one for partly AI-modified content. You may use them, but you don’t have to. In the Commission’s words: “The use of these icons does not establish legal compliance by itself.”

Breaking the Article 50 rules can lead to fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs, the lower of the two applies.

What to do now

  1. List every place where AI talks to or produces something for people outside your company: website chat, support bot, email assistant, product images, generated reports, social posts.
  2. Decide per system whether you are the provider or the deployer. Bought as-is: deployer. Built on an API and offered under your name: probably provider.
  3. For chatbots you provide, show a clear notice at the start of the conversation that the user is talking to an AI system. Don’t hide it in the terms.
  4. For generated images, audio and video you provide, ask your model vendor which marking it applies (watermark, metadata) and check that your own pipeline doesn’t strip it, for example when you resize, compress or re-encode a file.
  5. For published text, make sure a named person reviews it before it goes out, or label it.
  6. Write down what you decided and why. If a supervisor asks, you can show your reasoning.

This is not legal advice. Ask your lawyer or privacy officer how Article 50 applies to your systems.

How we approach this

Our blog engine writes articles with sources, and an editor approves every article before it goes live. That review step is good for quality, and it is also what the text exception in Article 50(4) asks for. For chatbots we put the notice in the interface itself, not in the terms, and log usage in a separate layer, as in the private AI workspace we built. If you want to check your own setup, see AI assistants and agents. For the AI Act rules on scoring people, which apply later, read automated decision-making and AI.

Sources

  1. Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the European Union, 24 July 2026
  2. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal of the European Union
  3. Code of Practice on Transparency of AI-generated Content, European Commission
  4. Guidelines on Transparency of AI-generated Content, European Commission
  5. EU icons for labelling AI-generated content, European Commission

More articles