New European Union guidance on labeling AI content
Vitaly Friedman unpacks rules that have caused a lot of confusion: the obligations are narrower and more sensible than reported, but since August 2, 2026, labeling AI content is a legal requirement for anyone serving European Union citizens. The piece separates the cases where a label is mandatory from those where it is not, explains why a sparkle icon in the interface is not enough, and sets out what the rules concretely require of a product with AI features. It bears directly on any client site with generated content or an assistant to talk to.
Article 50(4) of the European Union AI Act has been in force since August 2, 2026. If your site or product touches EU citizens and uses AI anywhere, the labeling obligation reaches you regardless of where the company is based. The scope rests on the same principle as the General Data Protection Regulation: it attaches to the user, not to the headquarters.
What must be labeled
- Deepfake content — an image, sound, or recording resembling real people or places and appearing convincingly authentic.
- A conversational assistant or agent, because the user must know they are not talking to a person.
- Text written wholly with AI assistance on matters of public interest, such as health, safety, the environment, the economy, finance, politics, science, and culture, if it has not passed human editorial review.
- Emotion recognition and biometric categorization tools.
What does not have to be labeled
The rule is narrower than the press coverage suggested, and the exceptions are explicit.
- Content that passed substantive human editorial review, with named responsibility.
- Minor assistive edits, such as spelling, grammar, formatting, image cropping, color correction, and translation.
- Closed business-to-business environments.
- Content a person meaningfully rewrote, rather than merely skimmed.
Why a sparkle icon is not enough
A sparkle icon on its own is ambiguous, because today it marks artificial intelligence, novelty, and recommendation alike, so it communicates nothing reliably.
A proper label is a clear icon paired with plain text. It stays visible when content is shared or downloaded, and it is available to assistive technology. The official European Union icon set is recommended.
Everything else is improper: a barely visible icon, a label buried in the page footer, a message that flashes or disappears, and an icon with no accompanying text.
Exactly who is bound
It binds both those who build and supply an AI system and those who use one in their own product. Embedding a licensed third-party tool does not release you.
That matters for agency work: when you embed someone else's conversational assistant for a client, the labeling obligation falls on the client as the system's user too.
Not only the European Union
China, California, South Korea, and India adopted similar rules through 2025 and 2026. The direction is the same everywhere, so it pays to solve this properly once rather than patch it market by market.
Sources
Related
How to label AI content under Article 50, and which part of it is not your job
Article 50 of the EU AI Act has applied since August 2, 2026, and it binds anyone serving people in the Union, wherever the server is. Most of the panic is about the machine-readable marking requirement, which for a site owner who calls somebody else's API is somebody else's obligation. Here is what is actually yours: a chatbot that says what it is, published text that either carries a name or carries a label, and a deepfake that admits it.
Evropska komisijaverified
Pew puts a number on it: 35% of pages published after ChatGPT show AI authorship
Pew Research ran nearly half a million pages from Common Crawl through a detector. Among pages published after November 2022, more than a third came back with significant signs of AI writing, and .com domains showed it at ten times the rate of .edu and .gov.
TechCrunchverified
Gemma passes a billion downloads and 100,000 derived models
On August 20, Google announced that the Gemma family of open models has passed a billion downloads. In two years the community has published more than 100,000 derived models, and the most recent Kaggle competition drew more than 1,600 entries. An Awesome Gemma repository launched with the announcement, meant as an index of vetted projects, fine-tuned models, tutorials, and tools. The figure matters to someone building sites too, because a model you run on your own server stops being exotic, which makes text processing without sending data to a third party workable.
Googleverified