Labelling AI Content: What the AI Act Requires
Most of the rules on labelling AI content have applied since 2 August 2026. On 2 December 2026 the last transition period ends. What do you have to do, and what is your supplier's job?
Labelling AI content: what is it about?
Do you use AI for copy, images, video or a chatbot on your website? Then Article 50 of the AI Act applies to you. That article regulates one thing: people must know when they are dealing with AI. The rules have applied since 2 August 2026. On 2 December 2026 the makers of existing generative tools must also start marking their output in a machine-readable way, and a new prohibition takes effect on the same day.
I see plenty of SMEs and marketing teams who think they now have to put 'made with AI' on every LinkedIn post. That is not right. Other companies think it does not affect them because their supplier takes care of it. That is not right either. The law draws a sharp line between whoever offers an AI system and whoever uses it, and the obligation lands on a different party depending on the situation.
The official text is in Article 50 of the AI Act, and the European Commission has published guidelines with it, full of examples. For the wider picture, from risk classes to deadlines, read the EU AI Act for SMEs. This article is not legal advice, but it gives you a practical route.
Provider or deployer: who has to do what?
Two roles in the AI Act matter here. The provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name. Think of the makers of ChatGPT, Midjourney or the chatbot software you buy in. The deployer uses an AI system under its own authority for its work. That is you, with your marketing team.
A few points from the guidelines that make a big difference in practice:
- Your employees are not separate deployers. The company is the deployer. If a freelancer works with AI under your direction, the company remains the deployer too.
- If you hire an agency and have no say in whether and how it uses AI, the agency is the deployer, not you.
- If you build a chatbot or generative tool yourself, even on someone else's model, and put it into service under your own name, you are the provider of that system.
- Purely personal use falls outside the deployer obligations. A freelancer using AI for client work is acting professionally and is therefore covered.
Chatbots: make clear it's AI
Article 50(1) requires the provider to design a system that talks directly to people so that those people know they are talking to AI, unless that is already obvious to a reasonably well-informed person. If you buy in a chatbot, that obligation formally sits with the supplier. If you build it, or have it built under your own name, it sits with you.
But it is your website and your customer. So always check that the notice is there and that you have not configured it away by accident. According to the guidelines, these work well:
- A first message saying the customer is talking to an AI assistant, possibly with a label that stays visible throughout the conversation.
- For a phone or voice assistant: saying out loud at the start that it is an AI assistant.
- For an agent that sends email: an AI label at the top of the message.
What is not enough
A sentence in your terms and conditions, a general notice such as 'Services on this website use AI', a vague word like 'assistant', or a human name with a photo that suggests a real employee. Does your customer service team use AI to draft replies, which an employee then checks and sends? Then there is no direct interaction with AI, and this obligation does not apply. That is how I prefer to set up customer service agents: the agent writes the draft, an employee checks it and presses send.
Deep fakes in marketing: when does it need a label?
Article 50(4) puts the obligation on you as the deployer: if you use AI to make images, audio or video that resemble existing people, objects, places or events and would falsely appear authentic, you must disclose that it was generated or manipulated by AI. That label has to be visible or audible. You cannot rely on the invisible marking added by your tool.
Not every AI image is a deep fake. The guidelines give examples you can use directly in marketing, whether you work with the tools in AI tools for marketing or with something else.
A deep fake, so label it
- An AI video with a realistic avatar of your own CEO.
- An AI influencer who resembles a well-known person, in an advert.
- An AI product image that makes the product look different, more appealing or better than it really is.
Not a deep fake
- A clearly fantastical scene, such as talking mice in a cheese advert.
- Your real product against an AI-generated background, as long as it does not mislead about how the product looks or works.
- Colour correction, removing a passer-by, noise reduction or resizing an image.
Creative work and older content
If the deep fake is part of an evidently artistic, creative, satirical or fictional work, the obligation is lighter: you disclose it in a way that does not spoil the work, for example in the credits. Content made before 2 August 2026 does not need to be labelled retroactively.
AI text on matters of public interest
The same Article 50(4) contains a second rule, for text. If you publish AI-generated or AI-edited text to inform the public on matters of public interest, you disclose that. Think of topics such as health, safety, the environment, politics and government, or economic developments.
Ordinary marketing copy and product descriptions are not covered, according to the guidelines, unless they contain claims about, for example, health, consumer safety or sustainability. A blog about your new collection needs no label. An AI-written article on the health effects of your product, or your sustainability report, does, unless the exception applies.
The exception: human review and editorial responsibility
You don't need a label if a person has reviewed the substance of the text and a person or organisation holds editorial responsibility for the publication. Both conditions must be met.
- Reviewing the substance means: someone who knows the subject checks the facts. A spell check or a quick glance is not enough.
- If AI changes the substance of the text after that review, the exception no longer applies.
- Make it easy to find who holds editorial responsibility, for example in an imprint or on a fixed page of your website.
Machine-readable marking: your supplier's job
Article 50(2) requires providers of generative AI to mark their output in a machine-readable way, for example with watermarks, metadata or cryptographic provenance data. That covers text, images, audio and video. New systems must do so from launch. For systems already on the market before 2 August 2026, the Digital Omnibus sets a deadline of 2 December 2026.
For you as a deployer, that means three things. Ask your suppliers how they mark output and whether they will be on time. Don't deliberately strip that marking from files before you publish them. And remember that the marking does not replace your own labelling duty for deep fakes and public-interest text: a person cannot see a watermark in the metadata.
If you build a generative application yourself and offer it under your own name, the marking obligation may sit with you. Have that checked before you go live. In June 2026 the Commission published a voluntary code of practice on marking and labelling AI-generated content, with practical steps for providers and deployers and a set of EU icons for labels.
New prohibition from 2 December 2026
On the same date a prohibition takes effect that the Digital Omnibus adds to Article 5: AI systems that create realistic intimate or sexual images of identifiable people without their explicit consent, and AI systems that create child sexual abuse material. For a normal marketing team little changes, but put it in your usage policy, so it is clear such tools never belong on a company device.
Fines and enforcement
Breaching the transparency rules can lead to a fine of up to 15 million euros or 3% of worldwide annual turnover. For SMEs, whichever is lower applies. Market surveillance authorities can investigate on their own initiative or act on a complaint, and anyone who suspects a breach may file one. The chance that a regulator knocks on a small company's door tomorrow over one LinkedIn image is small. The chance that a customer, competitor or journalist spots a misleading AI video is not.
Labelling is therefore mostly a governance question: who in your organisation decides when a label is needed? How to set that up as a board is covered in AI governance for SMEs.
Checklist: labelling AI content in your business
- List every place where AI talks to customers: chatbot, voice assistant, agents that send email.
- Check for each chatbot that the AI notice is visible at first contact, and not just in the terms and conditions.
- Agree with your marketing team what counts as a deep fake, using the examples above, and what your label looks like.
- Decide who determines whether an AI image, video or voice gets a label.
- Decide which texts cover matters of public interest, such as health, sustainability or safety, and who reviews their substance.
- State on your website who holds editorial responsibility for what you publish.
- Ask your AI suppliers how, and from when, they mark their output in a machine-readable way.
- Don't strip markings and provenance data from files before you publish them.
- If you build a chatbot or generative tool under your own name, have it checked whether you are the provider.
- Agree in your contracts with agencies and freelancers how they handle labels.
- Add the new Article 5 prohibition to your AI policy.
Key takeaways
- Article 50 has applied since 2 August 2026. On 2 December 2026 existing generative tools must also mark their output in a machine-readable way.
- Chatbot notices and machine-readable marking are provider obligations. If you build under your own name, that provider is you.
- You label deep fakes and AI text on matters of public interest yourself, visibly for people. A watermark in the metadata is not enough.
- Ordinary marketing copy and product descriptions need no label, except for claims about, for example, health or sustainability.
- Text whose substance a knowledgeable person has reviewed, under editorial responsibility, falls under the exception.
Your chatbot and AI tools set up properly?
With AI implementation I set up your AI tools and chatbot so the notices are right, your team knows when a label is needed and the agreements are in place. For a fixed monthly fee, with training included.
See AI implementation