Labelling AI-generated content: what the EU AI Act requires and how to prepare

The short answer

From 2 August 2026, Article 50 of the EU AI Act applies – the part of the regulation that sets transparency obligations towards the people who see your content. In practice it means four things. The chatbot on your website must say the customer is talking to AI. Realistic images, audio and video generated or altered by AI (so-called deepfakes) need a visible label. Generative tools must mark their outputs in a machine-readable format, which is their providers' duty. Text written with AI that has passed human editorial review needs no label. Below we translate these obligations into everyday marketing situations and show how to prepare your company without paralysing content production.

Where this comes from: the AI Act in three dates

The AI Act is Regulation EU 2024/1689, the world's first regulation covering the AI market as a whole. It entered into force on 1 August 2024, but its provisions were rolled out in stages: from 2 February 2025 the bans on the riskiest practices (such as manipulative social scoring systems), from 2 August 2025 the obligations for providers of general-purpose models, and from 2 August 2026 most of the remaining provisions, including the transparency obligations of Article 50 discussed here.

One detail of scope matters a lot: the regulation also covers companies outside the EU whenever their systems or content reach people in the Union. If your campaigns are produced by an agency or a tool from across the ocean, the obligations don't go away. You may recognise this staged roll-out mechanism from another EU regulation we've written about: the European Accessibility Act.

Who's responsible for the label: the tool provider or your company?

The AI Act splits the duties between the provider of an AI system and the deployer – the company that uses the tool. If your team generates graphics in Midjourney, texts in ChatGPT or voiceovers in ElevenLabs, your company is the deployer in the meaning of the regulation.

The split looks like this:

• The tool's provider is responsible for machine-readable marking: the outputs of a generative system have to be marked in a machine-readable format and detectable as artificially generated, for example through metadata or a watermark.
• Your company is responsible for the layer the audience sees: disclosing a deepfake, telling people they're talking to a chatbot and, in the narrow case described below, disclosing AI-generated text.

You can't build the watermark on the provider's behalf, but you do answer for which tools you use and what you publish. From August 2026, choosing tools with solid metadata support is part of legal compliance, not just convenience.

Which content needs a label? The four cases of Article 50

Chatbots and voice-based customer service. AI systems built for direct interaction with people must make it clear the person is dealing with AI, at the latest at first contact. There's an exception where it's obvious to a reasonably well-informed person, but obviousness is a weak thing to rely on: a short note in the chat window settles the matter. We describe how to design these experiences in our article on AI agents for business.

Deepfakes: realistic images, audio and video. A deepfake in the meaning of the regulation is content that resembles existing people, objects, places or events and could falsely appear authentic. Such material must carry a clear, distinguishable disclosure that it has been artificially generated or manipulated. For content of an artistic or satirical nature the obligation is lighter: the information can be given in a way that doesn't spoil the experience of the work.

Text on matters of public interest. The obligation to label AI-generated text applies only to publications that inform the public on matters of public interest, and even then it lapses when the content has passed human review and a specific person or company holds editorial responsibility. A typical company blog or newsletter with human editing is therefore not covered.

Emotion recognition and biometric categorisation. If a company uses such systems on people, it must inform them. In marketing communication this is rare, but we mention it for completeness because the deadline is the same: 2 August 2026.

Do you have to label AI images in your campaigns?

Three questions settle the answer. Does the material realistically depict people, places, objects or events? Could a viewer mistake it for an authentic photo or recording? Did AI generate or substantially alter the content, rather than merely assist standard editing? Three answers of "yes" mean a deepfake and a visible labelling obligation.

On one side of the line: a photorealistic "photo" of a product in the hands of a model who doesn't exist, a generated video of a "customer" praising your service, a voiceover imitating a real person. On the other: a vector illustration, an abstract key visual, an obvious 3D stylisation, plus ordinary retouching, colour correction or background removal, because such operations don't substantially alter the content. Treat borderline cases cautiously: a label costs one line of caption, while its absence can end in lost audience trust and proceedings before a supervisory authority.

Platforms have their own rules, independent of the AI Act

Meta adds "AI info" labels to content it detects or that users declare as AI-generated. YouTube requires creators to disclose realistic synthetic material. TikTok automatically labels content carrying C2PA (Content Credentials) metadata and can remove unlabelled synthetic content under its terms of service. The C2PA standard is provenance metadata attached by tools from Adobe and OpenAI, among others; Google marks the outputs of its models with SynthID.

The practical conclusion for you: you need a labelling process anyway, because the law and the platforms you publish on require it in parallel. Better to design it once, properly.

How to label in practice: three layers

The layer your audience sees. A short, unambiguous caption next to the material: "Image generated by AI", "Voice generated by AI", or a note in the chat window: "You're talking to an AI assistant". The disclosure should be visible at first contact with the content, not in the terms of service three clicks away.

The technical layer. Check whether the tools your team uses attach provenance metadata (Content Credentials or an equivalent) and whether your publishing process strips it out, because image compression and exports from graphic tools are the usual culprits. It's the providers' obligation, but verifying it is in your interest.

The process layer. Decide who on the team rules on labelling, what the exact wording is and where in the material you place it. Write these rules into your brand guidelines, next to the logo and typography rules. Add a clause to contracts with agencies and freelancers: a declaration of AI use and assigned responsibility for the label. While you're at it, take stock of content published earlier: the rules apply from August 2026, but unlabelled realistic AI material from before that date can still breach platform terms and your audience's trust.

Labelling rules belong in your brand guidelines

If your brand already has guidelines for AI tools, which we wrote about in our article on the AI-ready brand book, labelling rules are their natural chapter: when we caption material, what the caption says, where it sits in the composition, how the chatbot introduces itself. That way every person and every tool working for the brand applies the same standard, and you don't re-litigate the same doubts with every campaign. We prepare guidelines like these as part of visual identity projects.

Penalties and supervision

Breaching the transparency obligations carries an administrative fine of up to €15 million or up to 3% of total worldwide annual turnover, whichever is higher. The European Commission is working on guidelines and a code of practice detailing the methods of marking and disclosure, and EU member states are appointing their national supervisory authorities. This article is for information; in ambiguous cases, especially campaigns using people's likenesses, consult a lawyer.

Where to start: five steps

1. Take stock of AI content in your company. List where materials are made with AI: campaign graphics, texts, video, voiceovers, chatbots. Without this list, every next step is guesswork.
2. Sort materials by obligation. Realistic images, audio and video: visible label. Text with human editing: no obligation. Illustrations and obvious stylisations: no obligation.
3. Check your chatbots and voicebots. Every point where a customer talks to AI needs clear information at first contact.
4. Set a caption standard and add it to your brand guidelines. One wording, one placement, one decision process.
5. Tidy up vendor contracts. Agencies and freelancers declare AI use, and responsibility for the label is assigned before material goes out.

We work with AI daily: we implement it in companies, design with it and prepare brand guidelines that AI tools follow too. If you want to put content labelling in order in your organisation, tell us about your project and we'll advise where to start.