AI content labeling under the EU AI Act: what Article 50 means for GEO and website content
In short
Article 50 of the EU AI Act has applied in full since 2 August 2026, and most of the guidance written about it predates the Commission's own guidelines. Two things are now settled that marketing teams keep getting wrong: ordinary advertising copy sits outside the labeling duty, and AI drafts that pass substantive human review under named editorial responsibility are exempt altogether. What that leaves in scope is a fast-growing category of content, the editorial articles written to be cited in AI answers. It is also why scaile builds the exemption into every article it produces: your team approves each one, your own knowledge goes into it, and a named person stands behind it.
Simon WilhelmCo-founder
Simon is co-founder of scaile and works with enterprise teams on visibility in AI search, from the first audit to defensible Share of Voice numbers.
Key takeaways
Advertising copy and product descriptions are out of scope, with one carve-back: claims about health, consumer safety or sustainability pull the text back in.
Human editorial review is a real exemption, not a formality. Substantive review with fact-checking as the minimum, under a named responsible person, removes the labeling duty. Cursory sign-off does not qualify.
The Digital Omnibus did not postpone Article 50. It moved the high-risk deadlines and gave pre-existing systems until 2 December 2026 for machine-readable marking, which is a duty of AI providers rather than of the teams using their tools.
GEO content is the most exposed text type. Publishing to be cited in AI answers means writing exactly the explanatory, informative pieces Article 50 has in view.
What the exemption requires and what citable content requires are the same things: sourced claims, checked facts, original knowledge, and a named person standing behind it. Compliance and AI visibility come out of one process.
scaile generates articles that are factually correct, on brand and built to be quoted in AI answers, runs every one through the client's approval, and learns from each edit. Because the client's own knowledge goes in and a named person approves, our position is that these articles need no label. See how the engine works.
Since 2 August 2026, Article 50 of the EU AI Act applies in full, to every AI system in scope, regardless of when it reached the market. Deployers must disclose AI-generated or AI-manipulated text published to inform the public on matters of public interest.
Two exceptions do most of the work for marketing teams. Ordinary advertising copy and product descriptions sit outside the duty entirely, and content that has passed human editorial review is exempt from it.
For most teams that settles the question everywhere except one place: the editorial articles written to be cited in AI answers. Those pieces explain, inform and sit close to matters of public interest, which is exactly the territory Article 50 describes. Anyone running generative engine optimization seriously is producing the text type the rule is aimed at.
This article is not legal advice. It sets out the position as of August 2026 with sources. For a binding assessment of your own case, ask counsel.
The practical answer for a GEO programme is to build the human review into the workflow rather than bolt a label on afterwards, because the review is what removes the duty. That is how scaile is built: every article is researched from your knowledge base, fact-checked, and opened in an editor where your team checks the text and the sources and signs off before it publishes. Editorial responsibility sits with a named person on your side, which is the condition Article 50 actually names.
Which AI-generated content has to be labeled?
Only some of it. For text, Article 50(4) covers one specific case: AI-generated or AI-manipulated text published to inform the public on matters of public interest. Three categories decide what your team actually does.
Advertising and product descriptions carry no labeling duty. The Commission’s guidelines name AI-manipulated text in company advertising and product descriptions as an explicit example of content outside the scope. That exception has an exception: claims about health, consumer safety or sustainability remain covered. Green claims, supplement marketing and safety-relevant product copy therefore need more care than the rest of your catalogue.
Editorial content on matters of public interest is covered, unless reviewed. Blog posts and thought leadership on economic, health or environmental subjects can fall inside the duty when published purely AI-generated.
Deepfakes always require a label. AI-generated or manipulated images, video and audio that convincingly depict real people, places or events must be marked under Article 50(4), with no editorial exemption available. That includes synthetic influencers and executive avatars. It does not include ordinary photo retouching.
Why this hits GEO content hardest
Generative engine optimization works through content that answers a question completely, with evidence, in a way a reader can follow. That is what makes a passage quotable to an AI system. It is also the text type Article 50 means when it talks about informing the public on matters of public interest.
The consequence is an awkward ordering for most content programmes. The product page teams worry about is exempt. The guide on regulation, health, finance or sustainability, the one actually written to land in the AI answer, is not. So the labeling question gets settled at the core of a GEO programme rather than at its edge.
The useful part is that the exemption asks for what good GEO work involves anyway.
When does human review remove the labeling duty?
The text of Article 50(4) says the duty does not apply where AI-generated content “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”
The Commission’s guidelines turn that into two conditions, and both have teeth.
The review must be substantive, carried out by competent people, with fact-checking as the stated minimum. Superficial checks, purely procedural approval and automated review processes are named as insufficient.
Editorial responsibility must sit with an identifiable natural or legal person, findable somewhere like the site’s imprint. Naming someone does not substitute for the review; both conditions have to hold.
There is a third detail almost nobody has noticed. The exemption lapses if the content is substantively edited by AI after sign-off. Running one last AI pass over an approved article forfeits the protection the review just earned.
What a compliant publishing workflow for GEO content looks like
Draft with AI. Have a competent person review the substance and check the facts. Add your own knowledge. Record who approved it. Publish without further AI edits.
That is how scaile is built, and compliance was not the first reason for it. Four properties work together:
The platform generates articles that are correct, sound like you, and are built to be quoted. A research agent works from your knowledge base and current sources, so every statement traces back to one of them, and each fact is checked. The text is written against your Brand Kit, in your tone and with your positions. And it is structured the way AI systems need in order to lift a passage: the answer first, questions as headings, evidence at the sentence, an FAQ and structured data.
Every article goes through your approval. The piece opens in an editor where your team reads the text and the sources, comments, edits or regenerates sections, and signs off. Only then does it publish. That step is not a formality. It is where expertise and responsibility enter the text.
The engine learns from every edit. What your team corrects, cuts or rephrases feeds back. The next article lands closer to your language, knows your arguments better and needs fewer corrections. Over time you get a model of your own house rather than a generic text generator.
We always encourage adding your own content and ideas. A number only you have, a customer case, a judgement only your team can make: that is the part no competitor and no model can reproduce. For AI search it is the reason your page gets cited instead of somebody else’s. For the labeling question it is the point at which a text stops being purely AI-generated in the first place.
Taken together, those four give us our position, and we will state it plainly: articles that scaile generates, that pass our clients’ approval and carry their own knowledge do not require a label under Article 50. They satisfy both conditions of the exemption, substantive review and named editorial responsibility, and they go beyond it in substance. Your own case still belongs with your counsel, but the construction is unambiguous.
When a piece is later refreshed at the same URL, it passes the same approval again. That keeps the exemption intact, and it keeps the freshness AI systems use to decide which source is still worth citing.
EU AI Act deadlines: what applies when
| Date | What applies |
|---|---|
| 1 August 2024 | The AI Act enters into force |
| 2 February 2025 | Prohibitions and AI literacy obligations apply |
| 2 August 2025 | Obligations for providers of general-purpose AI models |
| 2 August 2026 | Article 50 applies in full, to all systems in scope |
| 2 December 2026 | Omnibus transition ends for machine-readable marking on pre-existing systems |
| 2 December 2027 and 2 August 2028 | Postponed high-risk deadlines |
Coverage through 2025 widely expected the Digital Omnibus to delay the transparency rules. It did not. Regulation (EU) 2026/1744, in force since 27 July 2026, leaves Article 50 untouched. What moved were the high-risk deadlines, plus a four-month transition for machine-readable marking that applies to AI providers rather than to the marketing teams using their tools.
Fines for breaching Article 50, and who enforces them in Germany
Breaches carry fines of up to €15 million or 3 percent of worldwide annual turnover, whichever is higher. For SMEs the lower of the two applies, so a company turning over €20 million faces a maximum of €600,000.
In Germany the Bundesnetzagentur is the competent authority, under national implementing law in force since 29 July 2026, with supervision formally assumed on 2 August. No enforcement is known so far, and the authority is visibly taking an advisory posture, with a service desk and published guidance rather than early cases.
The nearer risk is commercial. Competitors can act through German unfair-competition law, and whether Article 50 counts as a market-conduct rule under section 3a UWG has not yet been settled by the courts. Sections 5 and 5a, covering misleading conduct and withheld material information, remain available routes. In the short term, a competitor’s cease-and-desist letter is likelier than a regulator’s fine.
Does AI labeling apply on Instagram and LinkedIn?
The duty attaches to the content, not the channel. A deepfake needs its label on Instagram exactly as it would on your own site, and a human-reviewed expert article stays exempt on LinkedIn. Platform labels from Meta or TikTok can satisfy disclosure within the platform, but responsibility stays with whoever publishes.
What an AI content labeling template should contain
A liability disclaimer does not discharge Article 50. The duty is disclosure, not small print, and a label offers no protection against unfair-competition, personality-rights or copyright claims.
Two building blocks cover it. For content inside the scope, a clear label such as “This content was generated with AI,” visible at first interaction. No particular wording is prescribed. For content relying on the exemption, a named responsible person somewhere findable, plus an internal approval log recording who reviewed the piece, when, which facts were checked, the sign-off, and confirmation that no AI edits followed it.
Keeping that log pays for both questions at once. The same record that supports the exemption also documents that checked facts and named experts stand behind the piece, which is what separates a cited source from an ignored one in AI search.
The free AI Visibility Check shows where your brand currently stands in AI answers. To see how approval, original input and a learning engine work together on your own material, book a demo, or read how teams without an editorial function of their own set it up.
FAQ
When does AI content have to be labeled?
Text needs a label when it is AI-generated or manipulated and published to inform the public on matters of public interest, without human editorial review. Deepfakes depicting real people or events always need one.
Do AI texts on my own website need labels?
Ordinary ad copy and product descriptions do not, provided they carry no health, safety or sustainability claims. Editorial content on matters of public interest does, unless it has passed substantive human review under named editorial responsibility.
Is AI labeling mandatory on Instagram?
For content inside the scope, yes, regardless of channel. Platform labels can satisfy the disclosure there, but responsibility stays with the publisher.
Does Germany have its own AI labeling law?
The duty comes from the EU AI Act and applies directly. Germany has legislated enforcement, making the Bundesnetzagentur competent. A planned amendment to the state media treaty is not yet in force.
Is a wave of cease-and-desist letters coming?
Unresolved. Whether Article 50 is actionable under section 3a UWG has not been decided, though sections 5 and 5a create real exposure. In the near term a competitor letter is likelier than a fine.
Does naming someone in the imprint count as human review?
No. The exemption needs both substantive review, with fact-checking as the minimum, and named editorial responsibility. An imprint entry supplies only the second.
Do AI-written SEO and GEO articles need a label?
It depends on the topic and the process, not on the tool. Product and advertising copy is exempt. Guides and expert articles on matters of public interest fall in scope when published purely AI-generated. With substantive human review and named responsibility, the duty falls away there too.
Do articles generated by scaile need a label?
In our assessment, no. Every article is generated against checked sources and your knowledge base, passes your team’s approval, carries your own expertise and is published under named editorial responsibility. That satisfies both conditions of the Article 50 exemption. It is not a legal guarantee, and your own case belongs with your counsel.
Does an article lose the exemption when it is updated?
Only if the update is purely machine-made. Run it through the same review and approval as the original and the exemption holds. At scaile, every same-URL refresh passes the same gate.
Sources
Legal texts and official guidance were read directly. Verify quoted wording against EUR-Lex before relying on it in a filing.
- EU AI Act, Article 50. Transparency obligations for providers and deployers, including the human review exemption.
- EU AI Act, Article 99. Penalties, including the €15 million or 3 percent ceiling and the SME rule.
- Regulation (EU) 2024/1689 on EUR-Lex. The consolidated legal text.
- European Commission guidelines C(2026) 5054 of 20 July 2026 on the Article 50 transparency obligations. The out-of-scope examples, the review standard, the post-approval lapse and the platform-label point, published via the Commission’s AI regulatory pages.
- Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force 27 July 2026. The postponed high-risk deadlines and the December 2026 transition for machine-readable marking.
- Bundesnetzagentur. German market surveillance under national implementing law, supervision assumed 2 August 2026.
- Haufe on the labeling duty from August 2026. Practitioner summary of the human review exemption.
