The AI Act: the 2026-2027 timeline for a content publisher

The EU artificial intelligence regulation, regulation (EU) 2024/1689, entered into force on August 1, 2024, but its obligations apply in stages. For a content publisher, an agency or a marketing team publishing text and images produced with the help of an AI, the task is not to read all 113 articles but to know which ones apply, on which date, and what changed with the July 2026 revision. Here is the timeline, article by article.

The general timeline in article 113

Article 113 sets the dates of application. The regulation holds more deadlines than those below; three concern a content publisher:

• February 2, 2025: general provisions and prohibited practices (chapters I and II), which do not concern marketing content production.
• August 2, 2025: obligations on providers of general-purpose models, governance, penalties.
• August 2, 2026: the rest of the regulation, including article 50 on the transparency of generated content, the only one that touches a publisher directly.

The tracking site artificialintelligenceact.eu reproduces article 113 and its deadlines; the authoritative version remains the one on EUR-Lex.

Article 50: two obligations, two parties

Article 50 is short but it mixes obligations that fall on different people. They have to be separated.

First point: technical marking is not the publisher's job, it falls to the model provider, and its absence does not release the publisher from its own disclosure obligation, which is separate. Then, for text, the disclosure obligation covers a narrow field: content published "for the purpose of informing the public on matters of public interest". A product page, a commercial comparison, a how-to guide or a landing page fall outside it, and content that would fall inside still escapes disclosure where it has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for its publication. The regulation sets out two alternative routes here, not one: reviewing every article is not the only path, editorial control exercised over production opens the same right, on the cumulative condition that editorial responsibility is named. That is the subject of our article on the "AI-generated" label.

What the July 2026 omnibus changed

Regulation (EU) 2026/1744 of July 8, 2026, known as the digital AI omnibus, entered into force on July 27, 2026, days before the 2 August stage. On the part that concerns us, it did not change the substance of article 50, but it introduced a delay for machine-readable marking: systems already placed on the market before August 2, 2026 have until December 2, 2026 to comply with paragraph 2, through a new paragraph 4 in article 111 on systems already in service. That delay applies to providers, not to publishers. It explains why generated text does not yet all carry a statistical watermark while images already do at the main providers.

The Commission also published the final version of its guidelines on article 50 on July 20, 2026, and a voluntary code of practice on the transparency of generated content was finalized on June 10, 2026. Both documents pin down the notion of public interest, defined broadly (health, environment, consumer matters, fundamental rights, and economic, political, scientific or cultural questions of societal importance) and what substantial review means. They do not carry the force of the regulation, but they are what national authorities will lean on.

The penalties

Article 99 sets the fines. A breach of the article 50 transparency obligations falls under paragraph 4, which provides for up to €15 million or 3% of worldwide annual turnover, whichever is higher, with the cap set to the lower amount for SMEs. These are maxima, and national enforcement practice has yet to be written: in most member states, the designation of competent authorities and their supervisory doctrine were not settled at the date of this article. It would be unwise to read that as an absence of risk: in several member states, an unmet consumer information obligation can be attacked by competitors under unfair competition law, long before any regulator takes an interest. The German case is covered in our article on the UWG.

The timeline, summarized for a publisher

What a publisher should have done before the end of 2026

1. Qualify the production. For each content type, answer two questions: is this a deepfake within the meaning of the regulation, that is, an image, sound or video that appears authentic without being so? Is this text published to inform the public on a matter of public interest? Most marketing content answers no to both, but not all of it: a health, safety or sustainability claim can tip a text into scope.
2. Formalize the review circuit. For texts that would be in scope, human review or editorial control, with named editorial responsibility, removes the disclosure requirement. You still have to be able to prove it: who reviews, at which stage, who signs, and where the record is kept. A checked box is not enough; a documented process is.
3. Settle the visuals question upstream. Rather than labeling every image on every channel, with placement rules that differ from platform to platform, the safest route is a production rule: never make a visual that claims to show a real event, place or person.

All of this is editorial work, and it weighs less the more it is built into the production tool rather than recalled at each publication. It is also a credibility argument: a media property that can say who reviewed and signed each article has no need to shelter behind a disclosure.