Labelling AI content step by step. Six steps before 2 December

Labelling AI content means telling the audience that a piece of material was generated or altered by artificial intelligence. The duty comes from Article 50 of Regulation (EU) 2024/1689 and has applied since 2 August 2026, but systems placed on the market earlier have until 2 December 2026 to meet the machine-readable marking requirement.

Below are six steps, the rules for choosing a label, where to download the free European Commission icons, and the exemptions that are easy to miss.

Who does this apply to, and from when?

Article 50 splits the duties across two roles. The provider of a system is responsible for marking the output in a machine-readable format (paragraph 2) and for making sure users know they are talking to AI (paragraph 1). The deployer is responsible for human-visible disclosure of deepfakes and of text published on matters of public interest (paragraph 4). Most companies occupy both roles at once: deployer towards the tools they use, provider towards their own chatbot.

Step 1: Inventory - where does AI touch our content?

Start with a list of the places where artificial intelligence creates or alters what you publish.

Control questions: Do we run a chatbot, a voicebot or an assistant on the site? Does AI generate product descriptions, posts, images, subtitles or voice-over? Do we use tools for retouching, background replacement or face swapping? Does anyone on the team translate or edit text with a language model?

Without this list the remaining steps have nothing to attach to, and the items skipped most often are translations and marketing graphics.

Step 2: Role and legal basis - which paragraph applies to us?

For each entry on the list, settle two things at once: which role you are in and which paragraph creates the duty.

Control questions: Are we the provider or the deployer here? Is this synthetic content under paragraph 2, a deepfake under paragraph 4, or an interaction under paragraph 1? Is the content published to inform the public on a matter of public interest? Has it been through editorial review?

That last question is often decisive: Article 50(4) does not apply to text that has undergone editorial review and for which someone holds editorial responsibility.

Which label for what? Three rules

The European Commission has released a set of icons and assigned each of them a specific use. The rules are simple, and the difference between them comes down to how much of the content a human made:

• Base icon - where AI was involved in producing a deepfake (image, audio, video) or published text, but you are not resolving to what degree.
• "Fully AI-Generated" - where the entire piece came from a model, with no human-created elements.
• "Partially AI-Modified" - where existing human-made content was partly altered using AI. This is where most marketing work lands: retouching, background changes, rewriting, translation.

Where to download the icons, and how to use them?

Each icon comes in four colour variants: black, white, and both at 50% opacity. Available as SVG and PNG, zipped, free of charge and with no attribution required: EU icons for labelling AI-generated content.

Two things are worth remembering. Using these icons is voluntary - the Article 50 obligation is not. And placing an icon settles nothing on its own: the label has to be noticeable on first exposure to the content and embedded directly in the material, unless an equivalent alternative exists.

Step 3: Human-visible disclosure - is it obvious straight away?

The paragraph 4 duty is met when the audience learns about the AI involvement on contact with the content, not after opening the terms and conditions.

Control questions: Is the label visible on first display, without scrolling or expanding? Would someone outside the industry understand it? Can it still be read once the material is shared onward? Does its form avoid obscuring the content itself?

For artistic, satirical and fictional work the duty is lighter: it is enough to disclose that manipulation exists, in a way that does not hamper enjoyment of the work.

Step 4: Machine-readable marking - does the provider deliver it?

This is a separate duty and in most cases it is not yours. Article 50(2) sits with the provider of the system: the output must be marked in a machine-readable format, and the technical solution must be effective and robust.

Control questions: Do we know which of our providers mark their outputs? Is that confirmed in documentation rather than by a salesperson? Have we checked what happens to the marking after a file conversion or a screenshot? Have our providers told us about the 2 December 2026 deadline?

How such marking works in practice, and where it breaks, we showed on a concrete example in the article on Anthropic marking Claude's text.

Step 5: Chatbots - does the person know who they are talking to?

Article 50(1) requires that a person knows they are interacting with an AI system, unless this is obvious to a reasonably observant and circumspect person. The information has to arrive no later than at the first interaction.

Control questions: Does our chatbot say it is a bot before asking its first question? Does the voicebot say so at the start of the call? Is the message available to people using a screen reader? Does the assistant's name avoid suggesting a human?

A human-sounding name and a photorealistic avatar work directly against what this paragraph asks for.

Step 6: Evidence - can we show how we classified content?

Labelling is a chain of decisions, and a decision with no trace cannot be defended the first time someone asks. After this step the organisation should hold:

• labelling rules written into the AI usage policy;
• a list of places where AI creates or alters content, with the role and legal basis for each;
• confirmations from providers regarding machine-readable marking;
• a record of how borderline cases were classified, with the date and the person;
• a named person who resolves doubtful cases;
• a record of the rules being reviewed after tools or law change.

That set turns "we did try" into something you can actually show.

How we do it here

On every page of this site, above the footer, one sentence sits: part of the content is created with artificial intelligence, and a human checks each piece before publication. Next to it is a link to our AI usage terms, where we describe what exactly we use it for.

This is an example of the third rule in practice: the content is produced with human involvement and reviewed by a human, so it sits closer to "partially AI-modified" than to "fully AI-generated". It is not a legal requirement in our case, because we publish neither deepfakes nor text on matters of public interest within the meaning of paragraph 4. We do it because readers have a right to know.

On your own, or with a tool?

The six steps above can be taken with a spreadsheet. The difficulty shows up in upkeep: tools change, people join, and every new publishing channel is a new classification decision. That is why labelling rules belong next to the AI tool register and the policy, not in a document of their own.

That is how the AI TrustCERT platform works: the tool register, policies with acknowledgements and a record of decisions in one place. If you want to check which AI Act duties apply to you first, the AI Ready Check gives a preliminary reading in three minutes.

Summary

Labelling AI content comes down to six questions: where AI touches our content, which role and which paragraph of Article 50 applies, whether the disclosure is obvious straight away, whether the provider marks the output for machines, whether users know they are talking to a bot, and whether we can show how we classified things.

There are three rules for choosing a label: the base icon where AI was involved, "fully AI-generated" where no human contributed, "partially AI-modified" where human content was altered. The Commission icons are free and using them is voluntary - unlike the obligation itself.

For systems placed on the market before 2 August 2026, the deadline is 2 December 2026. Classifying a given piece of content stays with the organisation, and where cases are contested the call is worth confirming with a lawyer.

✦ This article was written with the help of artificial intelligence and reviewed before publication by the author, who takes editorial responsibility for it.