The AI Act's split calendar: what Article 50 actually requires on 2 August 2026

What happened

Two calendars now sit on top of each other, and only one of them is law.

The AI Act sets its general date of application at 2 August 2026 (Article 113). In November 2025 the Commission proposed a "Digital Omnibus" to defer parts of that calendar, and the co-legislators agreed a text: the European Parliament adopted it on 16 June 2026 (423 in favour, 57 against, 174 abstentions), the Council gave its final approval on 29 June 2026, and the act was signed on 8 July 2026.

But as of 23 July 2026 the Digital Omnibus has not been published in the Official Journal. The Parliament's own procedure file records the status as "procedure completed, awaiting publication in the Official Journal." An EU regulation enters into force on the third day after OJ publication, so until that day arrives the amended dates are not yet positive law and the AI Act's original calendar governs. Anyone planning solely against the deferred dates is planning against a text that is not yet in force.

Why it matters

The headline "2 August 2026 is delayed" is wrong, and acting on it is the expensive mistake.

What the Omnibus defers, per the Council and Parliament press releases, is the high-risk regime: stand-alone high-risk systems (Annex III) move to 2 December 2027, and high-risk systems embedded in regulated products (Annex I) move to 2 August 2028. The Omnibus also introduces a new prohibited practice — AI-generated non-consensual intimate imagery and CSAM, the "nudifier" ban — applying from 2 December 2026.

Article 50 is not in that deferral. Its transparency duties still apply from 2 August 2026: disclosure that a person is interacting with an AI system (50(1)); disclosure by deployers of emotion-recognition and biometric-categorisation systems (50(3)); and disclosure of deepfake and certain AI-generated public-interest content (50(4)). The single carve-out is Article 50(2) — the obligation to mark synthetic audio, image, video and text in a machine-readable format. For generative systems already on the market before 2 August 2026, that marking obligation is deferred to 2 December 2026.

This is worth stating precisely because primary sources disagree with some secondary summaries. At least one secondary source gave the embedded-high-risk (Annex I) date as both "2 August 2026" and "2 August 2028"; the primary text resolves it to 2 August 2028. It is a small reminder that on a split calendar, the date you cite should come from the regulation or an official press release, not from a paraphrase of one.

What changes in practice

For a provider of a system that talks to people or generates content, the operative date is 2 August 2026, not 2027 or 2028. Interaction disclosure and deepfake disclosure are due then. Only the machine-readable marking of a pre-existing generative system buys the extra four months to 2 December 2026.

The exposure is not theoretical. Article 50 breaches fall under the Article 99(4) penalty tier — up to EUR 15,000,000 or 3% of total worldwide annual turnover, whichever is higher. One nuance is often mis-stated: the penalty chapter (Chapter XII) has itself applied since 2 August 2025; what changes on 2 August 2026 is that Article 50 becomes an obligation capable of being breached, not that the fining power "switches on" that day.

Two adjacent dates belong on the same planning sheet. The nudifier/CSAM prohibition applies from 2 December 2026, and — on a separate track — the Machinery Regulation (EU) 2023/1230 applies from 20 January 2027, which matters for anyone placing AI-enabled machinery on the EU market. The Omnibus also adjusts how the AI Office's competence is drawn for general-purpose-model-based systems; the substance is set out in the Council press release, and the exact article numbering should be read from the Official Journal text once it appears rather than from pre-publication consolidations.

Open Questions

The decisive open item is procedural: when the Digital Omnibus is published in the Official Journal, and therefore on which day its deferrals actually take effect. Until then, two calendars coexist and only the original one binds. A widely repeated "publish by 30 July 2026" date is a commentator inference from the third-day entry-into-force rule, not a deadline stated in any EU primary source.

It also remains open how the marking obligation of Article 50(2) will interact with the detection and labelling codes of practice that the AI Office is to facilitate under Article 50(7): whether a common technical method emerges, or whether marking is satisfied in several incompatible ways. That is the question that decides whether "machine-readable" turns out to mean machine-verifiable across platforms, or merely machine-present within one.