Regulation (EU) 2026/1744
On 27 July 2026 the EU AI Act stopped being the text most guidance still describes. This page sets out every change the amending act made: what the position was, what it is now, and which date each change attaches to.
It delayed part of it, and expanded another part. The high-risk obligations moved out by 16 months for Annex III use cases and 12 months for Annex I products. Nothing already in force was rolled back, and two new prohibited practices were added, taking the Article 5 list from eight to ten.
That combination is why “the AI Act was delayed” is a misleading summary. An organisation reading it that way risks missing two bans that bite on 2 December 2026 and a marking deadline on the same date.
Two high-risk deadlines in Article 113 were deferred, and one date was created that does not appear in the adopted text at all.
| Obligation | Adopted text | As amended |
|---|---|---|
| Two new prohibitions and the legacy marking deadline | No such date | 2 December 2026 |
| Annex III high-risk obligations apply | 2 August 2026 | 2 December 2027 |
| Annex I product-safety high-risk obligations apply | 2 August 2027 | 2 August 2028 |
See every application date in context, including the dates the Omnibus left untouched.
The amending act reaches beyond the timetable. It rewrote an obligation, inserted two prohibitions, created an enforcement chapter and resolved two ambiguities.
2 August 2026
2 December 2027
A 16-month deferral for the eight Annex III use-case areas. The stated reason was that the infrastructure the deadline depends on, harmonised standards and designated notified bodies, was not ready.
2 August 2027
2 August 2028
A 12-month deferral, keeping the product-safety route eight months behind Annex III so sectoral conformity assessment has time to absorb it.
Not prohibited
Prohibited
A ninth ban, inserted rather than deferred. It sits in the €35 million / 7% penalty tier alongside the original eight prohibitions.
Not prohibited
Prohibited
A tenth ban, on the same date and in the same penalty tier. Anyone still working from a list of eight prohibitions is working from the superseded text.
No definition; any AI in a regulated product risked being caught
AI used solely for non-safety purposes is not a safety component
This narrows the Annex I route rather than deferring it. AI used solely for user assistance, performance optimisation, service efficiency, automation, convenience or quality control does not qualify as a safety component, unless its failure would endanger health and safety. A product needing third-party assessment only for non-health-safety risks, such as radio spectrum or electromagnetic interference, no longer meets the Article 6(1)(b) condition.
Providers and deployers shall ensure a sufficient level of AI literacy
Providers and deployers shall take measures to support AI literacy
Softened from a duty of result to a duty of effort. The obligation still binds every operator regardless of risk tier, and has applied since 2 February 2025, but it is now assessed on the measures taken rather than the level achieved.
No transitional window
Legacy systems must comply by 2 December 2026
Generative systems already on the market before 2 August 2026 got a transitional window to implement machine-readable marking of synthetic output. The deadline is set by a new transitional paragraph, Article 111(4); the duty it defers is the existing Article 50(2). The only amendment to Article 50 itself is to paragraph 7, on codes of practice.
No direct supervisory powers over these operators
Information requests, evaluations, remedial measures and fines, with periodic penalty payments
A new enforcement chapter. The AI Office can impose periodic penalty payments of up to 5% of average daily turnover for each day a breach continues, which on a long-running infringement can exceed the headline maximum fine.
Only SMEs, including start-ups, faced the lower figure
SMCs also face the lower of the fixed sum and the turnover percentage
A new category. SMEs already had the lower figure under Article 99(6); this extends the same treatment to small mid-cap enterprises, companies that have outgrown the SME thresholds. “SMC” is defined by new Article 3(14b) as a small mid-cap enterprise under point (2) of the Annex to Recommendation (EU) 2025/1099.
Only the narrow Article 10(5) basis
A broadened legal basis for bias detection and correction
Addresses a practical deadlock: testing a high-risk system for discriminatory outcomes can require the very special-category data the GDPR restricts.
Not present
New annex: a list of codes, categories and corresponding types of AI systems
Notification under Article 30 now uses a fixed taxonomy of AIA codes, so a notified body's designation states exactly which types of system it may assess. A body designated for one category cannot assess another.
The deferral buys time on the heaviest workstream and buys none at all on the nearest deadline.
The Digital Omnibus on AI is Regulation (EU) 2026/1744, an amending act that entered into force on 27 July 2026 and changed Regulation (EU) 2024/1689: the EU AI Act. It deferred the high-risk compliance deadlines, added two prohibited practices, softened the AI literacy duty, gave the AI Office direct enforcement powers, and clarified how penalties apply to SMEs.
It delayed part of it. The high-risk obligations moved: Annex III use cases from 2 August 2026 to 2 December 2027, and Annex I product-safety systems from 2 August 2027 to 2 August 2028. Nothing already in force was rolled back: the prohibitions, the AI literacy duty, the general-purpose AI model obligations and the Article 50 transparency rules all continue to apply, and the Omnibus added two new prohibitions.
Ten. Eight have been unlawful since 2 February 2025. The Digital Omnibus inserted Article 5(1)(ba) on non-consensual intimate imagery and Article 5(1)(bb) on child sexual abuse material, both applying from 2 December 2026. Material that still refers to eight prohibitions is describing the superseded text.
2 December 2027 for AI systems in the eight Annex III use-case areas, and 2 August 2028 for AI systems that are safety components of products covered by the Annex I harmonisation legislation. Both dates are later than the adopted text provided.
The duty was softened from ensuring a sufficient level of AI literacy to taking measures to support it. It still applies to every provider and deployer regardless of risk tier and has done since 2 February 2025, but compliance is now assessed on the measures taken rather than the level of literacy achieved.
Yes. Articles 75a to 75d, added by the Digital Omnibus, give the AI Office powers to request information, run evaluations, require remedial measures and impose fines on general-purpose AI model providers. It can also impose periodic penalty payments of up to 5% of average daily turnover for each day a breach continues.