EU AI Labels Become Law

If a customer-service bot told you this week that it was artificial intelligence, that was no longer a bit of corporate politeness. Since 2 August, it has been a legal requirement. Article 50 of the EU AI Act now requires AI systems and synthetic content to identify themselves for what they are (EU AI Act, Commission summary). The rule does not stop companies using AI. It tries to make the synthetic layer of the internet easier to see.

What actually has to happen

The new obligation lands in three practical places. A chatbot must tell you at the start of the conversation that it is AI. A realistic synthetic image, video or audio clip must carry a visible label. AI-generated text on matters of public interest, such as election explainers, public-health notices or policy summaries, must be disclosed as machine-made, unless a human editor has reviewed it and accepts editorial responsibility (Article 50 guidance).

There are two layers to the system. The first is visible: the tag on a deepfake video, or the notice beside a chatbot window. The second sits inside the file itself, through metadata that records where the content came from and can be read by detection tools even when the image or text looks ordinary (Code of Practice). It is close to the logic of food labelling: the consumer gets a clear notice, while regulators get a way to inspect what is underneath.

The duties are divided along the chain. OpenAI or another chatbot provider has to design the system so users know they are dealing with a machine, and build origin-tracking metadata into synthetic outputs. The bank, newsroom or political campaign using that tool has its own obligation: to tell people when they are looking at deepfakes or AI-written material on matters of public interest (Stibbe). A line hidden in terms and conditions will not do. The Commission says disclosure must come at the start of the first interaction, in clear and accessible language (AI Act FAQ).

Someone posting a personal AI meme is not the intended target. The rules apply to professional and commercial use, rather than purely personal activity (EU AI Act).

Complaints now have somewhere to go

The EU wrote the law, but member states have to make it work. Germany has given the clearest early signal: the Bundesnetzagentur, the federal network agency, became the default complaints office on 2 August, with citizens able to file reports free of charge through an online form (Bundesnetzagentur, German government). Ireland has created a national AI Office to coordinate the domestic response (Department of Enterprise Ireland). Other capitals are putting similar structures in place.

The penalties are large enough to be noticed. Transparency breaches can bring fines of up to €15 million or 3% of global annual turnover. That is not the €35 million or 7% ceiling sometimes cited in coverage; the higher tier is for prohibited AI practices, not labelling failures (EU AI Act, Commission enforcement note).

The label is a starting point, not a solution

The obvious problem is that bad actors will not queue up to comply. Content generated before the deadline does not have to be labelled retrospectively. A missing label also does not prove something is real, because the technical standards for origin-tracking metadata are still developing (Two Birds). Courts will have to settle the harder boundary questions: what counts as "obviously AI", and how prominent must a label be inside a particular app?

Nor is the whole AI Act arriving in one go. Generative AI systems already on the market have until 2 December 2026 to add machine-readable markings. Wider high-risk rules covering employment, education and credit scoring take effect in December 2027 for standalone systems, and in August 2028 for systems built into regulated products (AI Omnibus, RegDossier).

The system is incomplete and can be tampered with. Still, it gives users and regulators a legal handle for the first time. The test now is whether labels change behaviour, build trust, or become another bit of screen furniture people learn to ignore.