Do You Have to Label AI-Generated Content? What the EU AI Act Actually Requires
Answer Engine Summary
No, most AI-assisted marketing content does not need to be labelled under the EU AI Act. The text labelling duty in Article 50(4) applies only to content published to inform the public on matters of public interest — politics, public health, environmental protection, consumer safety and similar. Commercial marketing copy falls outside it. A second exemption removes the duty entirely for any text that has undergone genuine human review with someone bearing editorial responsibility. What does apply broadly: chatbots must announce themselves, and deepfakes must be visibly marked.
On 2 August 2026, the transparency obligations of the EU AI Act became applicable across all 27 member states. Within days, our inbox filled with the same message in three languages: does this mean everything we publish now needs an AI disclaimer?
It does not. But the reasons why are worth understanding properly, because the parts that do apply are the parts nobody is talking about.
Why Everyone Got This Wrong at the Same Time
Two separate stories collided in the same week and the headlines merged them into one.
The first: 2 August 2026 was originally the deadline for high-risk AI system obligations — the heavyweight compliance regime with risk management systems, conformity assessments and technical documentation. The Digital Omnibus package moved those to December 2027 and August 2028.
The second: the transparency obligations in Article 50 landed on exactly that date and were not postponed. Neither was the enforcement machinery behind them. So half the market read "the deadline moved" and stood down, while the other half read "the AI Act is here" and assumed every blog post now needs a warning label. Both are wrong, in opposite directions.
Four Obligations, Not One
Most of the confusion dissolves the moment you stop treating Article 50 as a single rule. It is four distinct duties, sitting on two different parties. The Act separates the provider — whoever develops an AI system and puts it on the market or into service under their own name — from the deployer, who uses it. If you publish AI-assisted content on your own website, you are a deployer.
Why Marketing Content Falls Outside
There are two independent reasons, and either one is sufficient on its own.
Reason one: the scope is narrow
The duty attaches to text that is published, informative in purpose, and about a matter of public interest. The Commission's guidance names the categories: politics, public administration, the administration of justice, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments capable of being the subject of public debate.
A landing page for a roof cleaning service is not that. A comparison article about self-storage pricing is not that. The provision exists to protect the information ecosystem, not to police commercial copy.
Reason two: human review removes the duty entirely
Text that has gone through human review or editorial control, where a natural or legal person holds editorial responsibility for it, is exempt. Even inside the public-interest categories.
The catch worth knowing: the Commission is explicit that superficial, purely formal or procedural checks do not qualify. Spellchecking is not editorial control. You need substantive assessment of the content and a named party with the authority to approve, amend or reject it. If your "review" is a glance before hitting publish, you are relying on an exemption you have not actually earned.
What About Images?
The visible labelling duty for images targets deepfakes, and the definition is narrower than most people assume. It requires two things simultaneously: the content resembles an existing — or convincingly existing — person, object, place or event, and it would appear authentic or truthful to a viewer.
A generic AI illustration on a blog post does not automatically meet that bar. An AI-generated "photo" of a completed project presented as a reference case very well might. The distinction is not the tool — it is whether a reader would read the image as documentary evidence.
Abstract illustration, stylised graphics, obviously synthetic hero imagery.
Photorealistic people, staged "team" or "customer" shots, product photography.
Synthetic before/after shots, fabricated reference sites, recognisable individuals.
The Duty Almost Everyone Overlooks
While the market argues about article footers, the clearest and least ambiguous obligation in the whole provision sits unattended on thousands of websites: if you run a chatbot, visitors must be told they are talking to an AI, clearly and at the point of first interaction.
There is no public-interest threshold, no human-review exemption and no interpretive grey zone. It takes an afternoon to fix. It is also the single most likely thing to be noticed by a regulator, a competitor or a customer's procurement department — because it is visible from the outside.
Enforcement and Penalties
Breaches of Article 50 carry fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher. Enforcement sits with national market surveillance authorities rather than centrally with the EU AI Office, which means the practical experience will vary between member states.
Proportionality is built in — the accompanying Code of Practice explicitly holds startups and SMEs to measures appropriate to their size and resources. Nobody is expecting a twelve-person company to run a compliance department. They are expecting you to have looked at it.
A Practical Checklist
Frequently Asked Questions
No. The labelling duty applies to text published to inform the public on matters of public interest, and it is lifted entirely where genuine human review and editorial responsibility exist. Ordinary commercial content falls outside both.
The high-risk obligations moved to December 2027 and August 2028. The Article 50 transparency obligations and the enforcement powers behind them were not postponed and applied from 2 August 2026.
Machine-readable marking under Article 50(2) is a provider obligation. If you are publishing content generated by someone else's system, it is not your duty. If you have built your own AI system and put it into service under your own name, the analysis changes and is worth checking with counsel.
The Act reaches organisations serving users in the EU regardless of where they are established. If your content is published to an EU audience, assume you are in scope.
The regulation did not ask whether a machine wrote the words. It asked whether a human took responsibility for them.
That was always the right question — the law just made it explicit.
This article is general information, not legal advice. Edge cases and contractual allocation of responsibility belong with a qualified lawyer. Primary sources: the European Commission's Article 50 FAQ, its transparency guidelines, and the Code of Practice on Transparency of AI-generated Content.