EU’s pioneering AI Act: staggered implementation raises questions and challenges

EU’s Pioneering AI Act: Staggered Implementation Raises Questions and Challenges

[Brussels] – The European Union has codified its much-anticipated AI Act under Regulation (EU) 2024/1689, marking a significant step towards regulating artificial intelligence. Officially coming into force on 1 August 2024, the AI Act will see its provisions gradually implemented over the next few years, impacting various operators in the AI sector. This phased approach, while intended to provide clarity and prevent market disruption, has raised important questions, particularly concerning the oversight of high-risk and limited-risk AI systems.

Phased Rollout: Key Dates and Provisions

The AI Act’s implementation will be revealed through key milestones:

• 2 February 2025: General provisions, including scope, definitions, and AI literacy, along with measures against prohibited practices like general-purpose social scoring and emotion recognition in workplaces.

• 2 August 2025: Obligations related to General Purpose AI (GPAI) models, the governance structure of the AI Act, and associated penalties.

• 2 August 2026: Full application date of most provisions, marking a critical juncture for high-risk AI systems and limited-risk systems.

• 31 December 2030: Specific obligations for AI systems integrated into large-scale IT systems before 2 August 2027.

Provisions for High-Risk AI Systems: A Potential Loophole?

Articles 111 and 113 of the AI Act introduce a complex matrix for high-risk AI systems. Systems placed on the market or put into service before 2 August 2026 are subject to compliance only if “significant changes” are made to their design post this date. This particular carve-out has garnered scrutiny, as it potentially allows non-compliant high-risk systems to remain in use unless substantial modifications are observed.

Defining “significant changes” remains an ambiguous task. Current interpretations link significant changes to unforeseen adjustments in an AI system’s logic, algorithms, or key design aspects. However, many argue that high-risk systems, designed for specific purposes, may not undergo such drastic transformations, leaving gaps in regulation.

Limited-Risk AI Systems: Uncertainties Abound

The AI Act’s approach to limited-risk AI systems, such as chatbots and AI systems generating synthetic content, is even more nebulous. Unlike high-risk systems, limited-risk systems see no express timelines or exemptions outlined in Articles 111 and 113. This raises the question of when compliance becomes a necessity for these systems placed on the market before 2 August 2026.

Some interpretations suggest transparency obligations under Article 50 should apply to all limited-risk systems from the 2026 date. Yet, the absence of explicit directives for older systems leaves room for different legal readings. Clarity on this front is anticipated from future EU Commission guidelines, which would hopefully streamline enforcement and compliance expectations.

Future Directions and Anticipated Guidelines

As the AI Act moves towards full applicability, the EU Commission’s forthcoming guidelines will play a crucial role in addressing existing ambiguities. The guidelines are expected to shed light on the interpretation of “significant changes” for high-risk systems and outline clear compliance requirements for limited-risk systems. An updated version of the Blue Guide on the implementation of EU product rules is also awaited, which could harmonise the regulatory approach across the board.

In summary, while the AI Act represents a landmark effort in AI regulation aimed at fostering innovation and ensuring human-centred technology, the phased implementation strategy, particularly with its exemptions, poses several questions. Stakeholders will look to the EU Commission for the necessary clarifications to navigate this new regulatory landscape effectively.

Source: Noah Wire Services