EU Artificial Intelligence Act amendments delay some deadlines and add new bans

A provisional Digital Omnibus on Artificial Intelligence would push back several EU Artificial Intelligence Act deadlines, refine how the law interacts with sector rules, and introduce new prohibited practices. The package also expands limited bias-testing allowances and strengthens centralized oversight for some high-impact systems.

Negotiators from the Council of the European Union, the European Parliament, and the European Commission reached a provisional agreement on 7 May 2026 on the Digital Omnibus on Artificial Intelligence, the first amendment package to the EU Artificial Intelligence Act since its adoption in June 2024. The package combines delayed implementation dates, narrower simplification measures, and a limited set of substantive policy changes. The most significant practical shift is a staggered deferral of certain compliance obligations that had raised concerns around testing, documentation, conformity assessment, and the readiness of supporting standards.

For high-risk Artificial Intelligence systems, the revised timetable follows a two-track structure. Annex III high-risk Artificial Intelligence systems would move from 2 August 2026 to 2 December 2027, while Annex I product-regulated systems would move from 2 August 2027 to 2 August 2028. Transparency obligations under Article 50(2) for systems generating or manipulating synthetic content and placed on the EU market or put into service before 2 August 2026 are postponed from 2 August 2026 to 2 December 2026, while systems placed on the market or put into service after 2 August 2026 must comply from that date. The obligation for Member States to establish at least one national Artificial Intelligence regulatory sandbox is postponed from 2 August 2026 to 2 August 2027. The package also gives standards bodies more time to prepare guidance that underpins compliance.

The amendments further refine the relationship between the Artificial Intelligence Act and sector-specific product rules. The Machinery Regulation ((EU) 2023/1230) is moved from Annex I Section A to Section B, shifting affected machinery away from a dual-compliance structure and toward sectoral regulation as the main framework. The Commission must adopt delegated acts amending Annex III of the Machinery Regulation to incorporate Artificial Intelligence-specific health and safety requirements by 2 August 2028. The definition of “safety component” is also narrowed so that systems used only for assistance, optimization, efficiency, automation, convenience, or quality control are excluded unless a failure or malfunction could endanger health or safety.

A notable policy addition is a new prohibition on systems used to generate or manipulate non-consensual intimate imagery and child sexual abuse material. That prohibition takes effect on 2 December 2026 and applies to both providers and deployers, though the trigger conditions differ. Providers face restrictions where prohibited output is an intended purpose or a reasonably foreseeable and reproducible outcome without significant technical modification, unless they have implemented adequate safety measures. Deployers are covered only when they intentionally use systems for those prohibited purposes, including through circumvention or misuse, while accidental generation is expressly excluded.

Other targeted changes carry operational and enforcement implications. A new Article 4a creates a narrow path to use special category data for bias detection and correction under strict necessity, security, access, retention, and documentation safeguards. The Artificial Intelligence Office gains exclusive supervisory authority over certain systems based on general-purpose Artificial Intelligence models and systems integrated into very large online platforms or search engines, with the Commission handling pre-market assessments for some high-risk systems subject to third-party conformity assessment. Information-sharing duties under Article 25 are made more specific, and breaches of Article 25(2) or (4) are added to the fining tier under Article 99(4), with penalties of up to 3% of worldwide annual turnover / €15 million. Core obligations including registration of high-risk systems and Artificial Intelligence literacy remain in place, although the literacy standard is softened from ensuring literacy to taking measures to support its development. Formal approval is expected in June, publication is expected in July, and the amendments would enter into force on the third day following publication in the Official Journal.

79

Impact Score

Qwen 3.5 raises concerns about censorship embedded in model weights

A technical analysis of Alibaba Cloud’s Qwen 3.5 points to political censorship circuits embedded directly in the model’s learned weights. The findings highlight operational, compliance, and product risks for startups building on third-party Artificial Intelligence models.

Laptop prices rise as memory shortages hit PCs

Laptop prices are climbing as memory makers redirect production toward data center demand driven by Artificial Intelligence. The squeeze is spreading beyond RAM to graphics memory and SSDs, raising costs across the PC market.

Artificial Intelligence models split on job disruption estimates

A new working paper finds that leading Artificial Intelligence models give sharply different answers when asked which jobs they are most likely to disrupt. The findings raise doubts about using model-generated exposure scores to guide labor policy or economic analysis.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.