The AI labelling duty: what applies since 2 August — and what the headlines leave out

In short: Article 50 of the EU AI Act has applied since 2 August 2026. Three things need labelling: chatbots, deepfakes, and AI-generated text published to inform the public on matters of public interest. Your ordinary product description, your proposal text and your internal meeting notes are not covered. Confuse the two and you either label too much or label the wrong thing.

What exactly has applied since 2 August 2026?

Article 50 of the AI Act took effect that day and governs four cases — not one. Anyone who only read “deepfakes must be labelled” knows a quarter of the rule.

Paragraph	Who is affected?	What is required?
Para. 1	Providers of systems that interact directly with people	Users must be able to tell they are talking to AI — unless it is obvious
Para. 2	Providers of generative AI systems	Mark outputs in a machine-readable format as artificially generated
Para. 3	Deployers of emotion recognition and biometric categorization	Inform the people exposed to the system
Para. 4	Deployers creating deepfakes or publishing AI text on matters of public interest	Disclose the artificial origin

The decisive distinction sits in the columns. A provider develops an AI system and places it on the market — OpenAI, Google, Anthropic. A deployer uses such a system in a professional capacity. The vast majority of companies are deployers, not providers. That makes paragraphs 1 and 2 primarily a question to put to their suppliers rather than to themselves.

Under the Commission’s guidance, the information is sufficient if it is provided clearly, unambiguously and accessibly no later than the time of the first interaction (Art. 50(5)).

Does the labelling duty really cover every marketing text?

No — and that is the most frequently skipped clause in the whole provision. Article 50(4) requires disclosure only for text published “in order to inform the public on matters of public interest”.

That is a real limitation, not a formality. A product description in an online shop does not inform the public on a matter of public interest. A proposal, a job ad, a newsletter about new opening hours, an internal set of meeting minutes: all outside. If you write those with AI, Article 50 does not require you to label them.

Inside the scope sits content with a public dimension: journalistic pieces, reporting on politics, health, safety, the environment or consumer protection, and contributions to societal debate. The boundary is fuzzy, and this is where the justified criticism of the rule begins: “matter of public interest” is not defined in the regulation itself. Companies publishing close to that line carry the uncertainty. In genuinely doubtful cases, labelling is the smaller risk — but only in doubtful cases, not as a blanket policy.

Do you have to watermark your AI content?

No, not if you are only using AI. The machine-readable marking duty in paragraph 2 is addressed explicitly to providers of AI systems, including general-purpose ones. Anyone using ChatGPT, Gemini or Midjourney professionally is a deployer and owes no watermark.

This confusion is circulating widely right now and is triggering projects nobody needs. What you should do instead is ask your suppliers. The regulation requires their technical solutions to be “effective, interoperable, robust and reliable”. For systems placed on the market before 2 August 2026, the Digital Omnibus pushes this obligation to 2 December 2026 — many tools will catch up by then.

For deployers, what remains is the visible disclosure under paragraph 4. That is an editorial matter, not a technical one.

When can you use AI without labelling?

In four situations — three of which matter in everyday business.

Assistive use. Where AI performs an assistive function for standard editing, or does not substantially alter the input data or their semantics, the duty falls away (Art. 50(2), third sentence). Spell check, grammar correction, formatting, translation without substantive change: no labelling.

Editorial control. For AI-generated text, disclosure is not required where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. The Commission’s guidelines are explicit: this means substantive engagement by qualified people, including fact-checking. A glance over the phrasing or an automated check is not enough. If the text is substantially altered by AI again after sign-off, the exemption falls away.

Important: this exemption covers text only. There is none for deepfakes in image, audio or video. An AI-generated advertising image showing a realistic-looking person stays subject to disclosure no matter how carefully it was reviewed.

Obviousness. The disclosure duty for direct interaction does not apply where a reasonably well-informed, observant and circumspect person would recognize the AI anyway.

Art and satire. For evidently artistic, creative, satirical or fictional works, the duty narrows to disclosing the presence of generated content in an appropriate manner that does not hamper the display or enjoyment of the work — a note in the credits rather than a banner across the image.

What does correct labelling look like?

There is no prescribed wording. The regulation requires the notice to be clear, unambiguous, accessible and given no later than the first interaction. On 10 June 2026 the Commission published a voluntary Code of Practice on marking and labelling AI-generated content, together with harmonized EU icons: “AI” as a general notice, “AI generated” for fully AI-generated content, and “AI modified” for partially edited media. Signing the Code is voluntary, but compliance can serve as evidence that the legal obligations are met.

Wordings that hold up in practice:

• Chatbot, on opening: “You are chatting with an AI assistant, not a person. For personal advice we’ll be glad to connect you.”
• Image or video: “Created with AI” — visible on the medium itself, not in a caption three paragraphs down.
• Partially edited material: “Edited with AI” plus one sentence on what was changed.
• Public-interest text without editorial review: “This article was created using AI.”
• Text under editorial responsibility: no mandatory label — but a transparency note plus a named responsible entity is the more credible option.

The most common implementation error is not the wording but the placement: a notice you only find after scrolling does not satisfy “no later than the time of the first interaction”.

What does a breach actually cost?

Up to 15 million euros or 3% of total worldwide annual turnover for the preceding financial year — whichever is higher (Art. 99 of the AI Act). The widespread shorthand “up to 15 million” is therefore only accurate for smaller companies. For a group with ten billion in turnover, the ceiling would be three hundred million.

Conversely, for small and medium-sized enterprises including start-ups, the lower of the two amounts applies. That takes the sharpest edge off the range for mid-sized companies, but changes nothing about the obligation itself.

A wave of fines is not realistic in the short term. Market surveillance is still being built up across member states, and authorities typically move first on deception cases rather than a missing notice on a chatbot. But concluding that the topic can be ignored misses the second lever: an unlabelled deepfake is also attackable under unfair competition law — and that needs no authority, just a competitor.

How do we handle this blog?

Transparently, because otherwise we would not be allowed to write it. Our blog posts are produced with AI support: research, structuring and the first draft run with AI assistance. Every post is then reviewed on the substance, every figure checked against the primary source, and sign-off is given by name at management level. Editorial responsibility rests with weooo GmbH.

That means the exemption in Article 50(4) applies — these posts carry no statutory labelling duty. We write it down anyway, because “do you work the way you recommend” is a fair question. And because the exemption does not hinge on the tool but on the review: publish AI text unchecked and you cannot invoke it.

What should you do this week?

Four steps that together take less than a day:

• Take stock. Where does AI touch the outside world in your company? A chatbot on the website, an AI phone assistant, generated images in campaigns, automatically written text. A list, nothing more.
• Classify instead of blanket-labelling. For each item, decide whether paragraph 1, paragraph 4 or nothing applies. Marketing texts without a public dimension drop out.
• Write down your editorial process. Who reviews, who signs off, who carries responsibility, and where can readers find that? Without this documentation you cannot invoke the text exemption in a dispute.
• Ask your suppliers. What machine-readable marking does your AI provider deliver, and from when? It is their duty, but your risk if they fall short.

One point that often gets lost: the AI literacy obligation for your staff under Article 4 of the regulation has applied since 2 February 2025. If you are looking at this topic anyway, handle both in one pass.

This is exactly where we work with companies: not plastering every text with a precautionary notice, but finding the handful of places where the duty genuinely bites — and leaving the rest alone. No buzzword bingo, potential over hype. What a cleanly scoped AI deployment looks like day to day is covered in Digital employees: what AI agents really deliver in mid-sized companies.

Conclusion

The labelling duty is not a muzzle on AI use. It is a rule about honesty in three clearly bounded places: tell people when they are talking to a machine. Tell them when an image or video is artificial. And tell them for text that informs the public on serious matters, unless a human has taken responsibility for it. Everything else stays as it was. Know the difference and you save yourself unnecessary notices — while being clean where it counts.

Unsure which of your AI applications fall under the duty? Talk to us — we’ll go through your use cases and sort what needs labelling from what does not.

This article provides a practice-oriented overview and does not replace legal advice in individual cases.

Sources

• Regulation (EU) 2024/1689 (AI Act), Article 50 — para. 1 information duty for direct interaction; para. 2 machine-readable marking by providers, exemption for assistive functions in standard editing; para. 3 emotion recognition and biometric categorization; para. 4 deepfakes and AI text on matters of public interest, exemption where human review and editorial responsibility exist, reduced duty for artistic and satirical works; para. 5 clear, unambiguous and accessible information no later than the first interaction. Applicable from 2 August 2026.
• Regulation (EU) 2024/1689, Article 99 — fines for breaches of Article 50 up to EUR 15,000,000 or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher; for SMEs including start-ups the lower amount applies. For comparison: prohibited practices under Article 5 up to EUR 35,000,000 or 7%.
• Regulation (EU) 2024/1689, Article 4 — AI literacy obligation for staff, applicable since 2 February 2025.
• European Commission, Code of Practice on marking and labelling of AI-generated content, published 10 June 2026 — voluntary code for providers and deployers of generative AI systems; harmonized EU icons “AI”, “AI generated” and “AI modified”; compliance can serve as evidence that obligations are met.
• European Commission guidelines on Article 50 — clarification of key concepts, requirements for human review (substantive engagement by qualified people including fact-checking; purely formal or automated checks are not sufficient), identifiability of the editorially responsible entity.
• Digital Omnibus on AI — postponement of deadlines for high-risk AI systems; the transparency obligations under Article 50 are explicitly excluded and apply unchanged from 2 August 2026. For providers whose generative systems were placed on the market before that date, machine-readable marking under Article 50(2) applies from 2 December 2026.