Provisional agreement reached on AI Omnibus package to amend EU AI Act

The European Parliament has announced that a provisional agreement has been reached on measures to amend the EU AI Act.

Key points of agreement

We don’t yet have the full text of the proposed amendments, but the European Parliament has announced that the key points of agreement are:

• High-risk AI regimes delayed
  • Annex III high-risk AI regime delayed until 2 December 2027.
  • Annex I high-risk regime (AI as safety components / products covered by EU sectoral legislation) delayed until 2 August 2028.
• Digital watermarking obligation delayed
  • Digital watermarking requirement in Art. 50(2) delayed from 2 August 2026 to 2 December 2026.
• New “nudification” prohibition
  • AI systems that create child sexual abuse material (CSAM) or non-consensual intimate images (NCII) will be prohibited from 2 December 2026.
• Softening of AI Act impact on sectoral legislation
  • AI systems as safety components in machinery will be governed by the Machinery Regulation, while all other regulated products will remain subject to the AI Act alongside their existing sectoral regimes.
• Personal data processing permitted for bias correction
  • Processing of personal data will be permitted where strictly necessary to detect and correct biases in AI systems, subject to proper safeguards.
• Enforcement Centralisation
  • Enforcement for general-purpose AI systems will be centralised within the EU AI Office.

Key points of uncertainty

Some important points won’t become clear until we see the proposed legislative text, including:

• High-risk grandfathering: Whether the Omnibus will extend the grandfathering period in Art. 111(2), such that the AI Act only applies to new systems placed on the market after the new delayed application dates.
• Digital watermarking delay: Whether the extension to the deadline for digital watermarking under Art. 50(2) applies only to existing AI systems placed on the market prior to 2 August 2026 or to all AI systems.
• Scope of nudifier ban: The precise scope of this prohibition is not yet clear.
• AI Office enforcement: The specific changes to centralise enforcement powers in the AI Office are not yet clear.
• AI literacy: There is no reference to softening the AI literacy requirement in Art. 4 in the European Parliament’s press release.

Next steps

The amendments still need to progress through several stages before becoming law, including formal adoption by the European Parliament and Council. However, we don’t anticipate any material changes or issues with these proposed amendments.

Impact

In light of this agreement, we think that organisations should:

• Prioritise compliance with the Art. 50 transparency obligations: Save for a short delay to the Art. 50(2) digital watermarking obligation, these obligations come into force on 2 August 2026 (the same date on which regulators under the AI Act obtain enforcement powers).
• Continue to focus on high-risk AI: Once the high-risk guidelines are published, organisations should concretely identify existing high-risk AI and implement processes to ensure future high-risk AI is identified early. They should also implement frameworks and processes to build compliance with the high-risk AI regime (as appropriate for providers and deployers), following technical standards where relevant.
• Prepare for enforcement: Regulators will have the power to enforce the prohibitions, transparency requirements and general-purpose AI (GPAI) model obligations (and potentially also AI literacy) from 2 August 2026. Organisations should take steps to prepare accordingly.