eu ai act 2026

The prohibited practices bit first, the AI literacy obligation followed in February 2025, the rules for large foundation models in August 2025. 2 August 2026 is the date for Article 50 — transparency. And Article 50 is not a single provision but a bundle of four obligations across four paragraphs.

The four paragraphs

Paragraph 1 requires that people can tell when they are talking to an AI rather than a human — the chatbot rule. Paragraph 2 obliges providers of generative systems to mark their outputs in machine-readable form. Paragraph 3 covers emotion recognition and biometric categorisation. And paragraph 4 — the most interesting one for creatives — requires deepfakes and AI-generated text on matters of public interest to be visibly labelled.

Provider or deployer?

The decisive dividing line: a provider is whoever develops an AI system and places it on the market under their own name. A deployer is whoever uses such a system professionally — you, your agency, your client. The machine-readable marking in paragraph 2 is the provider's job. The visible labelling in paragraph 4 is yours.

The AI-generated campaign image

A fully generated visual running as an ad or a social format, realistic enough to be taken for real — that is a case for labelling. A discreet but legible note is enough: “AI-generated” will do. The same goes for the voice clone in a podcast intro or the video in which real footage was translated with AI lip-sync. Rule of thumb: as soon as an AI creates something new out of a real face, a real voice or real gestures that was never recorded that way, it is a deepfake within the meaning of the regulation.

The AI-written text

Here lies the biggest relief — and the biggest misunderstanding. The obligation only bites when two things coincide: the text informs the public about a matter of public interest, and it has not undergone genuine editorial review. Classic advertising copy, product descriptions, a LinkedIn post about your own offering generally do not fall under it. They do not inform on questions of public interest; they sell.

The chatbot on the website

Running for three years without a notice? From August it needs one. Paragraph 1 requires users to be told at the first interaction that they are writing to an AI — unless that is obvious anyway. One sentence is enough. But it has to be there.

What happens if you ignore it

Breaches of the transparency obligations can be penalised with up to 15 million euros or 3 per cent of worldwide annual turnover — whichever is higher. For SMEs the lower figure applies, and the Omnibus provides for automatic reductions and, in part, a correction period. But “reduced” is not “irrelevant”. Add German unfair competition law: misleading AI content can attract cease-and-desist letters through competition rules. The risk is not only regulatory — it is also the competitor who notices the missing label.

The pragmatic roadmap

Panic is the worst adviser, ignorance the second worst. It starts with taking stock: where does AI-generated or AI-manipulated content actually arise? A simple AI inventory is the basis for everything else — you cannot label what you do not know about. Then clarify roles, then the workflow, and finally the contracts: freelancer clauses, CMS documentation, defined label positions for image, audio, video and text.

Orientation

The voluntary EU code of practice on transparency for AI-generated content, presented in June 2026, is not legally binding but is likely to become the central reference framework for regulators and courts. Follow it and you carry a lighter burden of justification in an inspection.

The honest core

This whole body of regulation turns on one simple distinction: did a human really think, check and take responsibility here — or did the machine run through on its own? The law forces the industry to answer that question visibly at last. For anyone who always held that AI-first means mediocrity-first, 2 August 2026 is not a bureaucratic scare. It is an invitation to put your own signature back on the label.