AI Labelling Obligations: What Art. 50 of the EU AI Act Requires Since 2 August 2026

When does AI have to be labelled?

Since 2 August 2026, Article 50 of the EU AI Act prescribes four cases: when people interact with an AI system (chatbots), when AI systems generate synthetic audio, image, video or text content (machine-readable marking), when emotion recognition or biometric categorisation is used, and for deepfakes and AI texts on matters of public interest (visible disclosure).

The EU AI Act (Regulation (EU) 2024/1689) has been in force since August 2024, but its obligations apply in stages. On 2 August 2026, the stage kicked in that affects most companies in daily business: the transparency obligations. Anyone operating a chatbot, generating product images or publishing AI content now needs to know what must be labelled and what does not. This guide walks through the duties one by one – including the exemptions that tend to get lost in the current excitement.

The AI Act timeline at a glance

The four labelling obligations under Article 50

1. Chatbots: people must know they are talking to AI

Under Art. 50(1), AI systems intended to interact directly with natural persons must be designed so that those persons realise they are interacting with AI. Exception: it is obvious anyway from the perspective of a reasonably well-informed, observant person. In practice this means the website chat needs a clear notice such as “You are chatting with an AI assistant” – at the latest at the time of the first interaction, not buried in the terms and conditions. A human name and a photo avatar without an AI notice are no longer a grey area since August 2026; they are a violation.

2. Synthetic content: machine-readable marking by the provider

Art. 50(2) obliges providers of generative AI systems to mark outputs in a machine-readable format as artificially generated or manipulated – via watermarks or metadata (for example the C2PA standard). This duty primarily hits the tool makers, i.e. OpenAI, Google or Adobe. It still matters for you as a user: stripping metadata from generated images to obscure their origin works against the purpose of the regulation – and where a visible disclosure duty applies (see below), that one is yours anyway.

3. Emotion recognition and biometric categorisation

Whoever operates systems for emotion recognition or biometric categorisation must inform the persons concerned under Art. 50(3). Most small and medium-sized companies rarely encounter this case – it becomes relevant, for instance, with AI-based voice analysis in call centres.

4. Deepfakes and AI texts on matters of public interest: visible disclosure

Art. 50(4) targets deployers – the companies using AI content. Two cases: first, deepfakes – artificially generated or manipulated image, audio or video content that resembles real persons, places or events and would falsely appear authentic. They must be visibly disclosed. Second, AI-generated text published to inform the public on matters of public interest. Here sits the decisive exemption: the labelling duty does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication.

Who has to label: provider or deployer?

The AI Act distinguishes between providers (who develop an AI system or market it under their own name) and deployers (who use it under their own authority). For labelling, the split looks like this:

How do AI texts have to be labelled?

The most important answer for companies first: a blog article, product description or newsletter that you create with AI support and then review and take responsibility for before publishing does not have to be labelled. The duty in Article 50(4) targets texts that inform the public on matters of public interest and are published without human review – fully automated news feeds, for example. Anyone exercising editorial responsibility falls outside the obligation.

If you do have to label (or want to voluntarily), a clear, well-visible notice on the content itself is enough, such as “This text was created with AI”. In July 2026 the European Commission published official EU icons for labelling AI-generated content – three symbols for “AI involved”, “fully AI-generated” and “partially AI-modified”. Using them is voluntary, but they create a recognisable standard format.

Is there a labelling obligation for AI images?

Yes, but a differentiated one. The machine-readable marking (invisible, in metadata or watermarks) is the tool provider's job. A visible label is owed by you as deployer only for deepfakes – i.e. when the image depicts real persons, places or events deceptively authentically. A recognisably illustrative AI visual, a stylised graphic or a generated background without reference to reality does not trigger a visible labelling duty. A photorealistic image of your CEO shaking hands with a celebrity at a meeting that never happened, however, does.

Nothing fundamentally different applies on social media: the AI Act attaches to the content, not the channel. On top of that, many platforms (Instagram, TikTok, YouTube) now require their own AI labels for realistic-looking generated media – violating those is a breach of platform rules, independent of the regulation.

Exemptions: what you do not have to label

• ▸AI-assisted texts, images and videos that a human has reviewed and takes editorial responsibility for – the normal case in content marketing.
• ▸AI as an auxiliary function: spell checking, translation, image optimisation or standard editing that does not substantially alter the content.
• ▸Obvious AI interaction: where everyone recognises that a machine is answering, the separate notice is dispensable – when in doubt, label anyway.
• ▸Art, satire and fiction: here a disclosure that does not spoil the work (e.g. in the credits) is sufficient.
• ▸Purely internal use without publication: an AI draft that stays internal reaches no public and no affected persons.

What fines are at stake?

Violations of the transparency obligations can be fined with up to 15 million euros or 3 percent of worldwide annual turnover – whichever is higher; for small and medium-sized enterprises, whichever is lower (Art. 99 AI Act). In Germany, the national implementation act has been in force since 29 July 2026; market surveillance is bundled at the Federal Network Agency (Bundesnetzagentur), which has set up a dedicated AI coordination centre including a service desk for companies.

Beyond fines, a second front is open: unfair competition law. A chatbot that poses as a human or makes false claims can be the subject of a formal warning independently of the AI Act – German courts attribute chatbot statements to the company as its own advertising. The background is covered in our article on AI chatbot liability.

Transition period: what applies to existing AI systems?

Generative AI systems that were already on the market before 2 August 2026 benefit from a four-month transition period until 2 December 2026 – but only for the machine-readable marking under paragraph 2. The chatbot notice, the deepfake disclosure and the information duty for emotion recognition apply since 2 August 2026 without grace period. Reading the transition period as a general moratorium is a mistake that invites an avoidable violation.

Checklist: implementing AI labelling in your company

• ▸Take inventory: where does AI interact with customers, where do you publish AI-generated content?
• ▸Check chat systems: AI notice visible in the chat window, before the first answer.
• ▸Include phone bots and voice agents – AI interaction must be disclosed there as well.
• ▸Deepfake check for image and video material: does it look deceptively real? Then label it visibly.
• ▸Document your editorial process: who reviews AI texts before publication? This secures the exemption for editorially controlled content.
• ▸Quiz your tool providers: do your generators mark outputs machine-readably? Do not strip metadata deliberately.
• ▸Respect platform rules: set the AI labels of Instagram, TikTok and YouTube in addition to the AI Act.
• ▸Assign ownership: one person is responsible for AI transparency and follows new guidance from the regulators and the EU Commission.

Our verdict

The AI labelling obligation is more precise than the headlines suggest. Not every AI text needs a label – editorial responsibility exempts you. But where people could be deceived, the regulation demands plain language: for the chatbot, the deepfake, the emotion recognition. Implementing these duties costs little and protects twice over – against fines, and against the loss of trust when customers discover the deception themselves. Transparency is turning from a nice-to-have into basic equipment of credible digital communication.