The short answer: Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved the EU AI Act's high-risk deadlines. Obligations for stand-alone high-risk systems in the Annex III areas (hiring, credit scoring, education, critical infrastructure and the rest) now apply from 2 December 2027 instead of 2 August 2026. High-risk AI built into products covered by Annex I legislation, such as toys and medical devices, follows on 2 August 2028. At the same time, the Omnibus added two new bans, on AI that generates non-consensual intimate imagery and on AI that generates child sexual abuse material, which apply from 2 December 2026. Generative systems already on the market before 2 August 2026 have until that same date to mark their outputs as AI-generated.

The AI Act was not paused and it was not repealed. Most of it is already in force. What changed is when the heaviest set of duties lands, and that extra time is best read as a testing window.

Every date below was checked against the consolidated text of the AI Act as at 27 July 2026, which the European Commission publishes on its AI Act Service Desk with each changed provision marked as new or amended. This is general information, not legal advice.

What the Digital Omnibus on AI is

The Commission tabled the proposal on 19 November 2025 as COM(2025) 836, procedure 2025/0359(COD), according to the European Parliament's research service. Its stated driver was that application of the AI Act had run into delays, particularly around the publication of harmonised standards and compliance tools for the high-risk requirements.

After that, the Council's AI timeline records its negotiating position on 13 March 2026, a provisional agreement with Parliament on 7 May 2026, and the Council's final green light on 29 June 2026. The provisional deal is where the fixed high-risk dates and the new ban on generating non-consensual sexual content and child sexual abuse material were settled.

The final act is Regulation (EU) 2026/1744 of 8 July 2026. Its EUR-Lex record gives 27 July 2026 as the date of entry into force. Besides the AI Act, Regulation (EU) 2024/1689, it also amends Regulations (EU) 2018/1139 and (EU) 2023/1230.

The new EU AI Act timeline, date by date

Here is how amended Article 113 and the transitional rules in Article 111 now read.

• 2 February 2025: Chapters I and II, including AI literacy and the original list of prohibited practices. Unchanged.
• 2 August 2025: notified body provisions, general-purpose AI model rules, governance and penalties (other than the fines for general-purpose AI model providers in Article 101). Unchanged.
• 27 July 2026: Articles 102 to 110, the amendments the AI Act makes to other EU laws, now apply from the Omnibus's own entry into force. This point is new.
• 2 August 2026: the AI Act's general date of application. Everything not carved out elsewhere applies from here, including the Article 50 transparency duties for systems placed on the market from this date. Unchanged.
• 2 December 2026: the two new prohibitions in Article 5(1)(ba) and (bb), plus the related paragraphs 1a and 1b. Also the deadline for generative systems placed on the market before 2 August 2026 to comply with Article 50(2). Both new.
• 2 August 2027: deadline for general-purpose AI models placed on the market before 2 August 2025 to comply. Unchanged.
• 2 December 2027: Chapter III, Sections 1 to 3, for high-risk systems classified under Article 6(2) and Annex III. Previously 2 August 2026.
• 2 August 2028: Chapter III, Sections 1 to 3, for high-risk systems classified under Article 6(1) and Annex I.
• 2 August 2030: deadline for providers and deployers of legacy high-risk systems intended for use by public authorities.

Which systems fall under the 2027 date

Annex III lists eight areas: biometrics, critical infrastructure, education and vocational training, employment and workers' management, access to essential private and public services and benefits, law enforcement, migration and border control, and the administration of justice and democratic processes. Concrete examples from the text include AI that decides admission to an educational institution, AI that screens and filters job applications, and AI that evaluates a person's creditworthiness (fraud detection is excluded).

If your product does one of those jobs and is not a safety component of a regulated product, 2 December 2027 is your date.

Which systems fall under the 2028 date

Annex I covers AI that is a product, or a safety component of a product, already regulated by EU harmonisation legislation and subject to third-party conformity assessment under that legislation. Toys and medical devices are among the sectors listed in Section A, and the Omnibus added the Machinery Regulation, (EU) 2023/1230, to Section B. Under Article 6(1), an AI system in this group is high-risk when both conditions hold: it is a safety component of the product or the product itself, and that product must undergo third-party conformity assessment under its sectoral law.

The two new prohibitions arriving on 2 December 2026

The Omnibus added two points to the list of banned practices in Article 5(1).

Point (ba) prohibits AI systems that generate or manipulate realistic images, video, audio or similar content depicting an identifiable person's intimate parts or sexually explicit activity, where that person has not given "freely-given, specific, informed, unambiguous and explicit consent." This is the provision commonly described as the nudifier ban.

Point (bb) prohibits AI systems that generate or manipulate child sexual abuse material within the meaning of Directive 2011/93/EU.

The new paragraph 1a matters most for builders. For providers, the ban is not limited to apps built for that purpose. It also covers systems where generating this content is a "reasonably foreseeable and reproducible outcome" of the system's design, training or capabilities, and where adequate technical safeguards against that output are missing, including safeguards against foreseeable misuse and for correcting misuse once reported. Deployers are caught only when they actually use a system to generate such content.

That wording turns a general-purpose image or video model's guardrails into a legal question. A provider that cannot show its safeguards hold up against foreseeable misuse is exposed. Breaching an Article 5 prohibition carries the Act's highest fines under Article 99(3): up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher.

The Article 50(2) marking deadline for existing generative AI

Article 50(2) requires providers of systems that generate synthetic audio, images, video or text to ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." The marking must be effective, interoperable, robust and reliable as far as technically feasible, with narrow exceptions for assistive editing functions and certain law enforcement uses.

New systems placed on the market from 2 August 2026 were caught on that date. The Omnibus added Article 111(4), which gives providers of systems already on the market before 2 August 2026 until 2 December 2026 to comply. The Council's timeline records that the co-legislators reduced this grace period from six months to three. Article 50 breaches fall under the Article 99(4) tier: up to EUR 15 million or 3% of worldwide annual turnover.

What did not change

It is easy to read the Omnibus as a general delay. It is not one. Several things are untouched or already live.

• The original prohibited practices have applied since 2 February 2025.
• General-purpose AI model obligations have applied since 2 August 2025.
• The high-risk requirements themselves were not deleted. Article 9 on risk management and Article 15 on accuracy, robustness and cybersecurity carry no new or amended markers in the consolidated text.
• Penalties for providers and deployers of high-risk systems under Article 99(4) remain up to EUR 15 million or 3% of worldwide annual turnover.

One legacy rule did shift. Under amended Article 111(2), high-risk systems already on the market before Chapter III starts to apply are only brought into scope if they undergo significant changes in design from that date, with a separate 2 August 2030 deadline for systems intended for public authorities. That makes the launch date of a product relative to December 2027 or August 2028 a planning variable in its own right.

For the full picture of the Act's risk tiers and how it sits alongside US state laws, see our complete guide to the EU AI Act and US AI laws.

Why the delay is runway for testing, not a reason to stop

The obligations being delayed are mostly about evidence, and evidence takes time to build.

Article 9(8) says testing of high-risk systems "shall be performed, as appropriate, at any time throughout the development process" and, in any event, before the system is placed on the market or put into service. It must be "carried out against prior defined metrics and probabilistic thresholds" suited to the system's intended purpose. Article 9(6) says testing has to ensure high-risk systems "perform consistently for their intended purpose."

Read together, those clauses mean a team cannot start testing on 1 December 2027. It needs defined metrics, thresholds it chose in advance, results gathered over development, and a record that shows consistency. A credit scoring model or hiring tool shipping in early 2028 needs that evidence built during 2027.

Article 15(5) is just as concrete. High-risk systems must be "resilient against attempts by unauthorised third parties to alter their use, outputs or performance," with measures against attacks such as "data poisoning" and "adversarial examples or model evasion." Robustness against a motivated attacker is something you show by attacking the system repeatedly and tracking whether fixes hold. It does not fit into a final pre-launch week.

A sensible use of the runway looks like this:

1. Classify every system now. Decide whether each AI feature is Annex III, Annex I, a transparency case under Article 50, or minimal risk. The answer sets your date.
2. Check the December 2026 items first. If you provide a generative model, the new Article 5 bans and the Article 50(2) marking deadline arrive in weeks, not years.
3. Write down your metrics and thresholds. Article 9(8) asks for them to be defined before testing. Doing this now gives you a baseline to measure against.
4. Test against adversarial behaviour, not just accuracy. Jailbreaks, prompt injection and model evasion are what Article 15(5) and the new Article 5(1a) safeguard language are about.
5. Retest after material changes. Models get updated and fine-tuned. A test record from a version you no longer ship carries little weight.
6. Map evidence to a recognised framework. Harmonised standards are still arriving, so many teams anchor their programme in the NIST AI RMF or ISO/IEC 42001 in the meantime. Our guide to the NIST AI RMF and ISO/IEC 42001 explains how those fit together.

Where independent testing fits

The AI Act's conformity assessment routes are set by law. Under Article 43(2), high-risk systems in Annex III points 2 to 8 follow the internal control procedure in Annex VI, with no notified body involved. Independent testing does not replace that. What it adds is a second set of results produced by someone with no stake in the outcome, which is useful when a buyer, insurer or regulator asks how a provider knows its system behaves as claimed.

RateAI evaluates customer-facing AI under the A.E.G.I.S.™ framework, which scores seven pillars: Hallucination Integrity, Guardrail Resistance, Data Handling and Privacy, Ethical Alignment, Autonomy and Control, Adversarial Robustness, and Explainability and Transparency. Guardrail Resistance and Adversarial Robustness, for example, cover the jailbreak, prompt injection and red-team testing that the Article 15(5) and Article 5(1a) language points toward. RateAI's Pro and Enterprise tiers include compliance mapping to the EU AI Act, NIST AI RMF and ISO 42001. Formal regulatory determinations remain with the relevant authorities. For background on how third-party AI certification works in general, read what AI safety certification is and how it works.

Frequently asked questions

When do EU AI Act high-risk obligations apply now?

For high-risk systems in the Annex III areas, from 2 December 2027. For high-risk AI covered by Annex I product legislation, such as toys and medical devices, from 2 August 2028.

Did the Digital Omnibus pause or repeal the EU AI Act?

No. It moved the high-risk dates and added new rules, but the prohibitions, general-purpose AI model rules and the 2 August 2026 general application date all stand.

What is Regulation (EU) 2026/1744?

It is the Digital Omnibus on AI, dated 8 July 2026, which amends the AI Act, Regulation (EU) 2024/1689. EUR-Lex lists 27 July 2026 as its date of entry into force.

What new AI practices are banned from 2 December 2026?

AI systems that generate or manipulate realistic intimate or sexually explicit content of an identifiable person without explicit consent, and AI systems that generate child sexual abuse material. Providers are caught when such output is the intended purpose or a reasonably foreseeable and reproducible outcome without adequate safeguards.

Do existing generative AI tools need to watermark their outputs?

Providers of generative systems placed on the market before 2 August 2026 must comply with the Article 50(2) machine-readable marking duty by 2 December 2026. Systems placed on the market from 2 August 2026 were covered from that date.

Does RateAI certification make a product compliant with the EU AI Act?

No. RateAI provides compliance mapping to the EU AI Act, NIST AI RMF and ISO 42001 in its Pro and Enterprise tiers, and formal regulatory determinations remain with the relevant authorities.

Plan your testing before the deadlines arrive

If you are working out which AI Act date applies to your product and what test evidence you will need before then, talk to the RateAI team about mapping your system to the EU AI Act and testing it under A.E.G.I.S.™.