The Deadline Moved. The Duty Didn’t.

Four days before it was due to bite, the EU moved its headline AI deadline — and a season of “Article 14 is now live” posts became wrong overnight.

On 27 July 2026 the Digital Omnibus on AI entered into force — the first amendment to the AI Act since 2024. Its most consequential move is a calendar one: the obligations for high-risk AI systems, the ones that carry the human-oversight duty, no longer apply from 2 August 2026.

In one paragraph: the Digital Omnibus defers the AI Act’s high-risk obligations for standalone Annex III systems — including Article 14 (oversight by design) and Article 26 (oversight, staffed) — from 2 August 2026 to 2 December 2027; AI embedded in regulated products under Annex I moves to 2 August 2028. Transparency, AI-literacy and general-purpose-AI duties keep their dates. The human-in-the-loop duty, though, was never only an AI Act duty.

What actually moved on 27 July

The Omnibus (reported as Regulation (EU) 2026/1744, in force 27 July 2026) replaces the Commission’s proposed conditional trigger with fixed dates. Standalone Annex III high-risk systems — recruitment, credit scoring, education, essential services, law-enforcement and border tools — now comply from 2 December 2027. AI embedded in regulated products under Annex I — medical devices, machinery, vehicles — moves to 2 August 2028. Article 14’s design duty and Article 26’s staffing duty travel with that package: for an Annex III system, the effective human overseer the law imagines is now a 2027 obligation, not a 2026 one.

What did not move matters as much as what did.

Transparency held its date. From 2 August 2026, Article 50 still requires disclosure that a person is dealing with an AI system and the marking of synthetic content, with a four-month grace to 2 December 2026 for the machine-readable watermarking duty on systems already on the market. AI-literacy (Article 4, in force since February 2025) stays, now softened to a duty to “take measures to support” literacy rather than guarantee it. General-purpose-AI model duties (since August 2025) stay. And a new Article 5 prohibition on AI-generated non-consensual intimate imagery and CSAM lands on 2 December 2026. The penalty architecture is unchanged: breach of the provider and deployer duties — Articles 16, 26 and 50 — sits in the €15 million / 3%-of-turnover tier, not the €35 million / 7% band reserved for the Article 5 prohibitions.

The duty the calendar can’t defer

Here is the part a deferral hides. The date that moved is one clock among three — and it is the only one that moved.

The second clock is the Product Liability Directive (2024/2853), which bites on every product placed on the EU market from 9 December 2026 — eleven days after the AI Act’s high-risk duty was originally due, and a full year before its new date. It treats software and AI systems as products, drops strict liability on them, and rewards a documented oversight record with a defence: a gap in that record can become a courtroom presumption of defect the operator must rebut (see The Product You Didn’t Know You Shipped). The Directive does not ask whether your AI Act clock has started.

The third clock is the oldest — general civil and professional liability. When an automated decision carries legal effect and a human signs it off, responsibility attaches to that human and their principal through contract, tort and professional duty, none of which wait for a compliance date. The “liability sponge” — the reviewer with a checkbox but no authority, no time and no record — is a design defect that harms today, whatever the AI Act says about 2027 (see The Human in the Loop Is a Legal Role). Designing that role — who holds the authority, on what record — is the work set out under Human in the Loop: the legal function.

Read together, the deferral is headroom, not a holiday. The EU moved the date because the standards and conformity infrastructure were not ready in time — not because effective human oversight stopped mattering. Sixteen months is roughly what it takes to build oversight that is load-bearing rather than decorative; an organisation that treats 2 December 2027 as a snooze button will arrive at it with the same paper reviewer it has now.

What to do with the sixteen months

General information on the AI Act (Regulation (EU) 2024/1689), the Digital Omnibus amending it, and the Product Liability Directive (Directive (EU) 2024/2853) — not legal advice, and no lawyer–client relationship is created. Applicability dates depend on a system’s classification (Annex I vs Annex III) and may change on further amendment; the Official-Journal reference for the Digital Omnibus should be confirmed against EUR-Lex. Any specific situation needs advice on its own facts.