Labelling AI Content: New Rules for 2026

Read FAQs

What happened: the Commission published the Code of Practice

The Code of Practice is a voluntary code of the European Commission that shows companies how to meet the transparency obligations under Article 50 of the AI Act. The Commission published the final wording of the code on 10 June 2026. Six independent experts worked on its preparation and more than 180 organizations took part.

Important: the code is voluntary. The obligations themselves, however, are not voluntary - they arise directly from Article 50 of the AI Act and will start to apply on 2 August 2026. The code is a “how-to guide”: whoever signs up to it and complies with it has practical certainty that they are meeting the obligations correctly.

The code also includes uniform EU icons for labelling AI content, which are meant to unify the visual form of labelling across Europe - much like cookie banners or nutrition labels once did.

Who does the labelling of AI content under Article 50 of the AI Act concern?

The key is to distinguish two roles. The same company can be in both at once.

Are you a provider of an AI system? Machine-readable labelling

If you develop or provide a generative AI system (typically AI startups, tool developers, SaaS with a genAI feature), you must ensure that all synthetic outputs - text, images, audio and video - carry a machine-readable label indicating that AI created them.

In practice this means:

• watermarking AI outputs (an invisible watermark embedded in the content),
• metadata according to open standards - typically C2PA. C2PA (Content Credentials) is an open standard for embedding verifiable metadata about the origin of digital content.
• ideally a combination of several techniques, because none is one hundred percent reliable.

Do you use AI in your company? Visible labelling only in three cases

If you only use generative AI (you are a so-called deployer - which is most companies: marketing, media, e-commerce…), you do not have to label everything. Visible labelling is mandatory only for:

1. A deepfake is, under the AI Act, image, audio or video content created or modified by AI that realistically imitates real persons, places or events and could mislead someone.
2. AI texts published with the aim of informing the public (e.g. news articles) - unless the text has undergone human editorial review and someone bears editorial responsibility for it.
3. Chatbots and interactive AI systems - the user must know that they are not communicating with a human.

In other words: a common marketing text that you have generated and that a colleague checks and approves does not have to be visibly labelled.

Exceptions: editing, art, satire

The AI Act also makes room for common sense. A milder regime, or none at all, applies to:

• assistive and standard editing - adjusting the brightness of a photo, grammar correction or an AI autocomplete do not turn content into “AI content”,
• artistic, creative and satirical works - these have a special, milder labelling regime so that the rules do not destroy creative work.

Are you developing an AI product? Watermarking, metadata and informing users must be built into the product before you put it on the market. We will give you a review of your AI product’s functional specification - we will go through it through the lens of the AI Act and tell you what to add, before it gets expensive to redo.

I want a specification review

From when does the labelling of AI content apply? Key dates

Watch out for that third row: the Digital Omnibus gave providers of older systems extra time for the technical implementation of watermarking and metadata. The deferral, however, concerns only machine-readable labelling for systems placed on the market before 2. 8. 2026 - visible labelling of deepfakes and chatbots applies from August for everyone.

🚨 NOTE:  The Digital Omnibus has not yet been finally approved, but we expect that this will happen soon.

What sanctions do you face for unlabelled AI content?

A breach of the transparency obligations under Article 50 can be costly: a fine of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher.

And alongside the fine there is reputational risk - an unlabelled deepfake or an AI article passed off as human work can damage customer trust faster than any fine.

Does it make sense to sign the Code of Practice?

For most companies yes, if they actively provide genAI or publish content with it. Signing the code brings:

• legal certainty - if you proceed in accordance with the code, you have a strong argument that you are meeting the obligations,
• lower compliance costs - you do not have to invent your own solution, the code refers to open standards and uniform EU icons,
• a signal to clients and partners that you take transparency seriously.

An alternative exists - you can also meet the obligations by your own route. But then you must prove yourself that your solution corresponds to the requirements of the AI Act. Details on signing can be found on the European Commission’s page on the code.

How to prepare for labelling AI content: a checklist

1. Map where you use AI. Which teams generate content? Do you operate a chatbot? Are you developing your own AI features?
2. Determine your role - provider, deployer, or both? The scope of your obligations follows from this.
3. Go through the published content. Does anything arise at your company that falls under deepfakes or AI texts for the public?
4. Set up an editorial process. Human review and editorial responsibility relieve you of the visible labelling obligation for texts.
5. Deal with the technology. If you are a provider, implement watermarking and C2PA metadata - by December 2026 for older systems, immediately for new ones.
6. Update internal directives on the use of AI and train the teams.
7. Consider signing the Code of Practice.