EU AI Act Omnibus Deal Reached: Postponed Deadlines, Watermarking Compromise, And The Nudificiation Prohibition

The Council of the European Union and the European Parliament reached a provisional agreement on the Digital Omnibus on AI in the early hours of 7 May 2026, after a nine-hour trilogue session under the Cypriot Council Presidency.

The text remains subject to formal endorsement by both institutions and to legal-linguistic revision before adoption, with publication in the Official Journal expected ahead of the original 2 August 2026 application date for the AI Act’s high-risk obligations. The agreement postpones the application of those high-risk obligations, narrows the grace period for the new transparency rules applicable to generative AI, and introduces a substantive new prohibition on the use of AI systems to generate non-consensual intimate imagery and child sexual abuse material. It does considerably less on the architectural questions concerning AI systems embedded in products already governed by EU sectoral product safety legislation, which had broken the previous trilogue at the end of April and which remain only partially resolved by the text agreed on 7 May.

This article sets out the principal elements of the agreement and their implications for organisations placing AI systems on the EU market or deploying AI within the scope of the AI Act.

Postponed high-risk application dates

The most consequential element of the agreement is the postponement of the application of the AI Act’s requirements for high-risk AI systems. As enacted, those requirements were due to apply from 2 August 2026 in respect of stand-alone AI systems falling within the use cases listed in Annex III, and from 2 August 2027 in respect of AI systems used as a safety component of, or themselves constituting, products falling within the EU sectoral harmonisation legislation listed in Annex I. The agreed text replaces those dates with fixed application dates of 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems. Both the Council and the Parliament had broadly converged on this calendar in their respective negotiating positions adopted in March 2026, and the principal function of the trilogue was to embed those dates in a final consolidated text before the original deadline arrived.

The rationale for the postponement is operational rather than political. The harmonised standards on which providers will principally rely to demonstrate conformity with the AI Act’s high-risk requirements are not expected to be available in their entirety until late 2026, and the designation of national competent authorities and notified bodies has lagged behind the legislative timetable in a number of Member States. Without an agreed postponement, providers and deployers would have faced obligations whose technical specifications had not yet been finalised and whose enforcement infrastructure was not yet in place.

AI in regulated products: a narrow carve-out for machinery

The element of the file that defined the failed trilogue of 28 April, and that has now been resolved on the narrowest possible basis, concerns the relationship between the AI Act and existing EU sectoral product safety legislation. AI systems within Section A of Annex I to the AI Act, which lists items including machinery, medical devices, in-vitro diagnostic medical devices, toys, lifts, and equipment for use in potentially explosive atmospheres, currently fall under a combined regime in which the sectoral conformity assessment applies. The AI Act’s high-risk requirements are layered on top. A coalition of Member States led by Germany, supported by industrial federations and elements of the European Parliament, had pressed for the Section A categories to be moved into Section B, which would have routed the AI-related requirements through the relevant sectoral instruments alone.

The agreement reached on 7 May makes that move only in respect of machinery. Medical devices, in-vitro diagnostic medical devices, toys and the other Section A categories remain within the AI Act’s horizontal regime. The Council and the Parliament have instead instructed the Commission to specify, by delegated act and within a period running to August 2027, how the substantive requirements of the AI Act for high-risk AI systems should integrate with the conformity assessment procedures already provided for under the Medical Device Regulation, the In-Vitro Diagnostic Medical Devices Regulation, the Toy Safety Directive and the other sectoral instruments listed in Section A. The architectural question that defined the dispute has therefore been deferred to instruments that do not yet exist, and on which the Commission will be expected to consult before adoption.

Article 50 watermarking: a four-month grace period

For providers of generative AI systems, the practically significant element of the agreement is the narrowing of the proposed grace period under Article 50(2) of the AI Act, which requires that synthetic outputs of generative AI systems (audio, image, video and text) be marked in a machine-readable format and detectable as artificially generated or manipulated. The Commission’s November 2025 proposal would have postponed the application of that obligation by six months. The Parliament’s negotiating mandate, adopted on 26 March 2026, sought a three-month postponement. The compromise reached in the trilogue sets the deadline at four months, meaning that providers placing generative AI systems on the EU market will be required to comply with Article 50(2) marking obligations from 2 December 2026.

The operational benchmark for compliance with Article 50, the AI Office’s Code of Practice on Transparency of AI-Generated Content, is itself still in draft. A first version was published in December 2025, a further draft is expected during May, and the final text is not anticipated before June 2026. Industry submissions during the trilogue process had argued that the Commission’s six-month grace period was already insufficient, given the evolving state of the underlying guidance and had pressed for a 12-month postponement. That position did not prevail. Providers will have a matter of weeks between the publication of the final Code of Practice and the application of the marking obligation to deploy the multi-layered marking, metadata embedding, fingerprinting, and detector tooling contemplated by the draft Code, together with the supporting governance and contractual architecture.

New prohibition on AI-generated non-consensual intimate imagery and CSAM

The substantive prohibition that the European Parliament had made non-negotiable in trilogue, and that constitutes the principal new restriction introduced by the omnibus, is a ban on the use of AI systems to generate non-consensual sexual or intimate content and child sexual abuse material. The provision will apply from 2 December 2026.

The drafting reaches beyond purpose-built nudifier tools to systems whose functionality renders such misuse reasonably foreseeable. The standard is therefore one of foreseeability rather than intent or labelled purpose. In practical terms, providers and deployers of generative image and video systems will be expected to evidence design-stage and deployment-stage controls, including content filtering, prompt-level safeguards, model fine-tuning constraints and incident response procedures, capable of preventing or mitigating reasonably foreseeable misuse for the prohibited purposes.

Other elements of the agreement

The agreed text also reinstates an obligation for providers to register high-risk AI systems in the EU database, even if the provider considers the system to be exempt from high-risk classification. It reinstates the strict necessity standard for the processing of special categories of personal data for bias detection and correction. The deadline for competent authorities to establish national AI regulatory sandboxes is postponed. Provisions on the supervisory remit of the AI Office in respect of AI systems built on general-purpose AI models have been refined, with a number of carve-outs reserving competence to national authorities in defined areas, including law enforcement, border management, judicial authorities and financial supervision. The detailed content of the AI Office’s enforcement powers will be set out in a Commission implementing act. The General-Purpose AI obligations under Articles 50 to 55 of the AI Act, which have been in force since 2 August 2025, are not amended by the omnibus and continue to apply on their existing terms.

Implications for providers and deployers

For providers of stand-alone Annex III high-risk AI systems, the agreement extends the runway to substantive compliance by sixteen months, from 2 August 2026 to 2 December 2027. For providers of AI systems embedded in regulated products under Annex I, the extension is 12 months from the original 2 August 2027 date. In each case, the relief is genuine, but the underlying compliance work remains unchanged: technical documentation, post-market monitoring obligations, conformity assessment engagement with notified bodies, AI risk management system design, and the integration of AI-specific controls into existing quality management systems must continue to be built out against the new dates.

For providers of generative AI systems, the position is more pressing. Article 50(2) marking obligations will apply from 2 December 2026, and the Code of Practice defining compliance is still in progress. Providers will need to have multi-layered marking and detection capabilities, along with supporting governance, contractual, and supplier flow-down architecture, in production by that date. Providers of any AI system whose functionality could foreseeably be used to generate non-consensual intimate imagery or content sexualising minors will additionally need to have misuse mitigation in place by the same date, supported by documented foreseeability assessment and design-stage and deployment-stage controls.

For deployers, including in-house users of AI in employment, education, biometric identification, credit scoring, critical infrastructure, law enforcement, justice and migration, the extended timeline applies in full. The work of AI inventory, classification under Annex III, fundamental rights impact assessment under Article 27, where applicable, governance design, human oversight protocols, transparency notices and (where relevant) worker information and consultation should now be planned against the 2 December 2027 date for Annex III deployments and the 2 August 2028 date for deployments in regulated product contexts. Deployers using generative AI in content production for the EU market should plan for the 2 December 2026 date for the Article 50 labelling and watermarking obligations, and align internal labelling workflows with the requirements of the final Code of Practice once published.

Conclusion

The agreement reached on 7 May closes a politically contested file before the original 2 August 2026 application date for the AI Act’s high-risk obligations, and it does so on terms that broadly track the convergence reached between the Council and the Parliament in March. The principal substantive new restriction it introduces is the prohibition on the generation of non-consensual intimate imagery and child sexual abuse material, applying from 2 December 2026. The principal areas of substantive uncertainty that remain are the integration architecture between the AI Act and EU sectoral product safety legislation, which is to be specified by delegated acts during the period running to August 2027, and the operational content of the Article 50 transparency obligations once the Code of Practice on Transparency of AI-Generated Content is finalised. For organisations placing AI systems on the EU market or deploying AI within the scope of the AI Act, the postponement should be treated as an extension of the timeline for delivery of the substantive obligations, rather than as a reduction of those obligations.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.