What Actually Takes Effect on 2 August 2026 (updated after the Digital Omnibus)
CORRECTION — July 2026. This brief was first published before the EU’s Digital Omnibus on AI amended the AI Act. Since then, the answer to its central question has materially changed: several of the obligations originally due on 2 August 2026 have been deferred, one new prohibition has been added, and the transparency obligations remain on their original date. This version corrects the brief in full. What follows is the current picture; the original is superseded.
On 2 August 2026, the obligations that actually take effect are, for the most part, the transparency ones — the duty to tell people they are dealing with AI. The heavier obligations most readers associate with that date — the high-risk requirements that would let a person challenge or stop an automated decision about them — have been deferred. That is the single most important correction to make if you have been planning against the AI Act’s original calendar: the disclosure layer arrived on schedule; the decision-accountability layer did not.
What changed since this brief was first published?
The European Commission proposed the Digital Omnibus on AI in November 2025 as a simplification package. After a failed negotiation in April 2026, the institutions reached a provisional political agreement on 7 May 2026; the European Parliament endorsed it on 16 June and the Council gave final approval on 29 June 2026, with publication in the Official Journal following shortly after. The result amends the AI Act just weeks before its central milestone — which is why any brief written before mid-2026 now states the 2 August picture incorrectly.
What actually takes effect on 2 August 2026 now?
The Article 50 transparency obligations remain largely on their original 2 August 2026 date. In practice this means the duties to inform people when they are interacting with an AI system, when content is AI-generated, and when they are subject to emotion-recognition or biometric-categorisation systems. If your programme treated August 2026 as fully relieved by the Omnibus, that is the most common and most dangerous misread — the transparency clock did not move.
Also already in force, and unaffected: the Article 5 prohibitions on the most harmful practices (applicable since February 2025), and the general-purpose AI obligations under Articles 51–55 (applicable since August 2025). None of these were touched by the Omnibus.
What was deferred, and to when?
The obligations that slipped are, notably, the ones carrying decision-accountability:
- Annex III high-risk systems (stand-alone, use-based — including AI used in employment, credit, and essential services): deferred from 2 August 2026 to 2 December 2027 — a sixteen-month move.
- Annex I high-risk systems (AI embedded in regulated products): deferred from 2 August 2027 to 2 August 2028.
- Watermarking / machine-readable marking of synthetic content under Article 50(2), for systems already on the market: deferred from 2 August 2026 to 2 December 2026. Systems placed on the market after 2 August 2026 must comply from the date they are placed.
- National regulatory sandboxes: the obligation on Member States to establish at least one is deferred to 2 August 2027.
What was added?
The Omnibus did not only relax. It introduced a new Article 5 prohibition, effective 2 December 2026, on AI systems used to generate or manipulate non-consensual intimate imagery (“nudifier” applications) and child sexual abuse material. This sits in the highest enforcement tier — the prohibited-practices ceiling of up to €35 million or 7% of global annual turnover — and, unlike the deferred high-risk obligations, it arrives on a near-term date.
What this tells us
Read together, the pattern is specific and worth naming plainly: the transparency obligations held their date; the decision-accountability obligations were deferred by more than a year. Making the presence of AI visible to people arrived on schedule; making automated decisions challengeable by them did not. This is the asymmetry the Institute tracks as the synthocracy pattern, and it is treated as a scenario-level indicator in the Futures strand (see Signposts, Dial 6 — The Deadline).
Two cautions keep this honest. First, a deferral is not a dismantling: the Act’s risk-based architecture remains intact, and the same package that postponed the high-risk dates also added a new prohibition on its original timeline. Second, the deferrals reflect a real operational problem — the standards and conformity-assessment infrastructure needed to make high-risk obligations workable had not materialised in time, and an unworkable binding deadline is its own failure. The observation is not “delay equals capture.” It is narrower: when the difficulty consistently falls on the accountability side and never on the disclosure side, the asymmetry is the finding.
FAQ
Has the EU delayed the AI Act?
Partly. Some obligations were deferred — the high-risk requirements (to December 2027 and August 2028) and watermarking (to December 2026). Others, including the Article 50 transparency duties, remain on their original 2 August 2026 date. “The EU delayed the AI Act” is an imprecise summary of a staggered change.
If I deploy AI in employment or credit decisions, when do the high-risk obligations bind me?
For stand-alone Annex III systems, the current date is 2 December 2027, following the Omnibus deferral — moved from the original 2 August 2026.
Does the deferral mean I can stop preparing?
No. The underlying work — inventorying and classifying AI systems — does not get easier with time, and existing sectoral law (product liability, data protection, anti-discrimination) continues to apply to AI-caused harm regardless of the AI Act’s dates.
When do the new nudifier/CSAM prohibitions take effect?
2 December 2026, in the highest enforcement tier.
