Article 50 applies from today. The AI labelling duty, read practically.

Today is 2 August 2026, and with it Article 50 of the AI Act, formally Regulation (EU) 2024/1689, applies. It is the article with the transparency duties: anyone who generates or uses artificial content has to say so under certain conditions. Of all the parts of the regulation, this is the one that actually touches most small studios. The prohibitions applied to very few, the high-risk duties hit only specific product categories. But images, sound and text from generative tools: that is everyday work.

The article is shorter and milder than its reputation suggests. It rewards a close reading, because two opposite mistakes circulate around it. One: ignoring it all, on the assumption that one is too small for Brussels. The other: disclosing more than the regulation asks for out of anticipatory thoroughness, up to and including one’s own toolchain.

Four situations, two roles

Article 50 governs four situations, and for all four it is decisive which role you are in. The regulation distinguishes the provider, who develops an AI system and places it on the market, from the deployer, who uses a system under their own authority. Anyone using a third-party generative tool is a deployer. Anyone building an AI system and making it available to others is a provider.

First: whoever operates an AI system that interacts directly with people, say a chatbot, has to make sure those people know they are talking to a machine. Unless that is obvious from the circumstances anyway.

Second: providers of systems that generate synthetic audio, image, video or text content have to mark the outputs in a machine-readable format so they are detectable as artificially generated. That is a technical duty at tool level. It falls on whoever makes the tool, not on whoever works with it.

Third: deployers of emotion recognition or biometric categorisation have to inform the persons concerned. Irrelevant for most studios, mentioned for completeness.

Fourth, and practically the most important case: deployers who generate or disseminate a deepfake have to disclose that the content was artificially generated or manipulated. Deepfake here means content that deceptively resembles real persons, objects, places or events and could falsely appear authentic.

What does not fall under it

The exceptions matter as much as the duties, because they limit the scope honestly.

The providers’ marking duty does not apply where the AI system performs merely an assistive function for standard editing or does not substantially alter the input. A tool that fixes typos or denoises a photo will, on that basis, as a rule not create synthetic content that needs marking.

For artistic, creative, satirical or fictional works, the deepfake disclosure is limited to an appropriate form that does not hamper the display or enjoyment of the work. So the regulation does not demand a warning stamped across the image. A notice in a findable place, say in the accompanying text, the colophon or the imprint, should be enough.

And for text, the disclosure duty applies only where it informs the public on matters of public interest, and even then not where a human has reviewed the text and someone holds editorial responsibility. That is the clause protecting the normal editorial case: whoever writes, checks and stands behind the result with their name does not label every bit of research assistance.

Labelling does not mean opening the workshop

Here sits the biggest practical misunderstanding. Article 50 requires the information that content was artificially generated or manipulated. Nowhere does it require what with. Which models, which tools, which workflows, which prompts: all of that remains a trade secret, and there are good reasons to keep it one. A tool list in the imprint is not transparency, it is an assembly manual for imitators, and it goes stale faster than anyone can maintain it.

On this reading, a broad, honest wording is enough. For instance: some imagery is AI-generated and curated by hand, it depicts no real persons or events. That says what the regulation aims at, without giving away more than necessary.

Implementation in an afternoon

For a small studio the implementation is manageable if approached as an inventory.

First the stocktake: where does synthetic content arise in your own operation, and where is it published? Websites, app content, covers, illustrations, audio. For each place, the role question: provider or deployer?

Then the notice in the right place. For imagery on a website, a clear section in the imprint or image credits is enough. Where a single work could appear deceptively real, the notice belongs closer to the work. In apps it belongs where the AI works, not in a PDF nobody finds. And whoever offers a generative system themselves needs the machine-readable marking of outputs, for which open standards such as C2PA are establishing themselves.

Finally, the stance on edge cases: the information has to be clear and distinguishable at the first encounter at the latest, accessible to people with disabilities, and it must not hide behind jargon. When in doubt, label more broadly, not in more detail.

What is at stake

Breaches of Article 50 can be fined under Article 99(4) with up to 15 million euros or 3 percent of worldwide annual turnover. For small and medium-sized enterprises, the lower of the two amounts applies. The Commission has announced guidelines on application, and at EU level a code of practice is taking shape that is meant to make the detectability of artificial content concrete. Whoever starts today with an honest, broad notice will have little to sharpen once those codes arrive.

The real reason for labelling, though, is not found in Article 99. Saying openly where machines helped takes the embarrassment out of the subject and the distrust out of the audience. Transparency is not penalty avoidance. It is the cheapest form of credibility there is.

Frequently asked

Does every AI-generated image need its own caption?

No. Article 50(2) requires the provider of the AI system to mark outputs in a machine-readable format. Anyone who merely uses AI-generated images on their own website is a deployer and only has to disclose deepfakes, meaning content that deceptively resembles real persons, places or events. For recognisably illustrative imagery, a general, findable notice should suffice, for instance in the imprint or image credits. What matters is that the information is clear and accessible.

Does the labelling duty apply to text?

Only in one narrow case: when AI-generated or AI-manipulated text is published to inform the public on matters of public interest. And even then the duty falls away if the text has undergone human review and a natural or legal person holds editorial responsibility. Anyone who writes or answers for their own texts and uses AI as a tool regularly has no disclosure duty here under the wording of the regulation.

Do the AI tools used have to be disclosed?

No. Article 50 requires the information that content was artificially generated or manipulated, not what with. Tools, models, prompts and workflows are trade secrets and may remain so. A broad, honest wording covers this under the wording of the regulation.

What are the penalties for breaching Article 50?

Article 99(4) of the regulation provides for fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises including start-ups, Article 99(6) applies the lower of the two amounts. The duties apply since 2 August 2026.