EU institutions have reached a provisional political agreement on the Digital Omnibus on Artificial Intelligence, introducing targeted changes to the EU Artificial Intelligence Act ahead of the 2 August 2026 compliance milestone. The most significant shift is a delay to the high-risk regime. High-risk obligations for stand-alone Annex III systems are deferred to 2 December 2027, while obligations for Artificial Intelligence embedded in regulated products under Annex I move to 2 August 2028. Formal adoption and publication in the Official Journal are still required before the changes take legal effect.
Despite those delays, 2 August 2026 remains an active compliance date. The Article 50 transparency obligations for Artificial Intelligence systems largely remain on the original schedule. A limited exception would give systems already placed on the market before 2 August 2026 a four-month grace period until 2 December 2026 for the Article 50(2) watermarking requirement. Other transparency duties, including disclosure when users are interacting with Artificial Intelligence systems, would still proceed as scheduled from 2 August 2026.
The agreement also adds a new Article 5 prohibition targeting Artificial Intelligence systems that generate or manipulate non-consensual intimate images, video, audio or similar material, or child sexual abuse material. For providers, the ban would reach not only systems intended for that use, but also systems where such output is a reasonably foreseeable and reproducible outcome without significant technical modification and where adequate safeguards are missing. The new prohibition is subject to a transitional period until 2 December 2026.
Additional changes would narrow double regulation for Artificial Intelligence embedded in products already covered by sectoral safety laws. The Commission would be able to limit specific Artificial Intelligence Act requirements where equivalent obligations already exist, and systems covered by the Machinery Regulation would be largely exempted. The deal also narrows the definition of “safety component,” so some systems used only for user assistance, performance optimization, service efficiency, automation, convenience, or quality control would not automatically be treated as high-risk unless failure or malfunctioning would endanger health and safety.
Other provisions broaden the legal basis for using special-category personal data for bias detection and correction across all Artificial Intelligence systems and general-purpose Artificial Intelligence models, subject to a strict necessity standard. The deadline for member states to establish regulatory sandboxes is postponed to 2 August 2027, and the EU Artificial Intelligence Office would gain clarified supervisory authority and stronger enforcement powers. The Article 4 literacy duty, in force since 2 February 2025, would be softened so providers and deployers must support the development of Artificial Intelligence literacy among staff rather than guarantee a specific level.
