EU eases parts of its Artificial Intelligence Act

The EU has agreed targeted changes to its landmark Artificial Intelligence Act, delaying some deadlines, narrowing parts of the high-risk category, and cutting overlapping compliance requirements. The package also adds a ban on tools that generate non-consensual sexually explicit images.

The EU has agreed to simplify parts of its Artificial Intelligence Act, its flagship law for regulating Artificial Intelligence. The changes were agreed on May 7 through an “Artificial Intelligence omnibus”, a package of targeted amendments tied to a broader digital simplification drive. The aim is to reduce red tape, address overlapping rules, and give businesses more room to comply while keeping the law’s core risk-based structure in place.

The biggest immediate change is timing. High-risk Artificial Intelligence systems under Annex 3 of the Artificial Intelligence Act, covering employment, education, and health insurance, now face a compliance deadline of December 2, 2027, delayed from summer 2026. Artificial Intelligence embedded in physical products like medical devices or industrial machinery gets more time, with obligations delayed until August 2028. The definition of high-risk has also been narrowed, so only systems whose failure would create genuine health or safety risks face the toughest requirements. Tools that mainly assist users or optimise performance no longer automatically fall under the full regime.

The revised framework also trims overlaps with other EU laws. Where sector-specific legislation already regulates Artificial Intelligence functions in areas such as aviation, medical devices, or financial services, companies will no longer face parallel assessments under both systems. Machinery has been removed entirely from the Artificial Intelligence Act and will instead be governed by sector-specific regulation. That shift was welcomed by companies including Siemens and ASML, but it also raised concerns that the EU could start fragmenting its regulatory approach.

For businesses, especially SMEs and small mid-cap firms, the package offers lighter administrative burdens. Simplified technical documentation, extended deadlines, and broader access to regulatory sandboxes are meant to make compliance easier. The changes are intended to be proportional, with a small company using an off-the-shelf chatbot facing less scrutiny than a company selling high-risk Artificial Intelligence for hiring decisions. Even so, compliance costs remain, and fines are still possible for violations.

The package also introduces a notable new restriction. A ban on Artificial Intelligence tools that generate non-consensual sexually explicit images, including deepfakes, takes effect December 2. The broader debate now centers on whether the EU can keep pace with rapid technological change through formal legislation alone, or whether the Commission and the Artificial Intelligence Office will need to rely more heavily on guidance, codes of conduct, and enforcement. Formal approval by EU governments and the European Parliament is expected in the coming months.

79

Impact Score

UC Berkeley law tightens Artificial Intelligence rules without banning it

UC Berkeley Law is adopting a stricter Artificial Intelligence policy that limits student use in core legal work while preserving space for specialized courses and instructor discretion. Faculty behind the change say the aim is to protect foundational lawyering skills as generative tools become more capable and more common in practice.

Artificial Intelligence food leaders see disruption moving at uneven speeds

Food technology leaders say Artificial Intelligence is reshaping the food system quickly, but adoption remains uneven across functions such as strategy, compliance, and research and development. Companies already see immediate returns in some areas, while others expect broader impact to take longer.

Lisa su pitches AMD as China’s alternative to NVIDIA

AMD used its Shanghai developer event to position China as central to its roadmap and to court developers looking for an alternative to NVIDIA’s CUDA ecosystem. The strategy focuses less on headline chip specs and more on migration support, open-source tools, and long-term bets on the next wave of Artificial Intelligence applications.

DeepWeb-Bench tests limits of deep research models

DeepWeb-Bench is positioned as a tougher benchmark for evaluating whether frontier language models can handle real deep research tasks beyond existing tests. Results point to derivation and calibration, rather than retrieval, as the main weaknesses in current Artificial Intelligence systems.

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.