EU AI Act Draft Offers Lenient Rules for Big Tech AI Models

As the May deadline approaches for finalizing guidelines under the EU AI Act, authorities have released a third and likely final draft of the Code of Practice for general purpose AI (GPAI) providers. Developed since last year, this updated version comes alongside a dedicated platform to improve accessibility. Stakeholders can submit written feedback until March 30, 2025.

Key Aspects of the Draft Code

The EU's risk-based AI framework imposes specific requirements on major AI developers regarding transparency, copyright compliance, and risk management. This Code helps GPAI creators understand obligations and avoid potential sanctions reaching up to 3% of global revenue for violations.

Streamlined Structure

The latest revision presents a more focused framework with refined commitments, incorporating feedback from December's second draft. Additional consultations will further refine the guidelines before final adoption, with experts aiming for enhanced clarity in the definitive version.

The document organizes requirements into several critical areas:

• Core commitments for GPAI providers
• Detailed transparency protocols
• Copyright compliance measures
• Safety and security obligations (specifically for high-risk GPAISR models)

Transparency Measures

The guidance includes sample documentation templates to help downstream implementers access necessary compliance information. However, contentious language around copyright compliance continues to generate debate.

Copyright Concerns

The draft employs qualifiers like "reasonable measures" regarding web crawling practices and copyright protection, potentially allowing significant leeway. Earlier stronger provisions about direct complaint resolution have been softened, now only requiring designated points of contact.

Notably, the text permits GPAI providers to disregard complaints deemed "manifestly unfounded or excessive" - raising concerns about automated copyright detection systems being ignored.

Safety Requirements

The EU AI Act's safety evaluations currently apply only to exceptionally powerful models (exceeding 10^25 FLOPs). Recent feedback has led to further narrowing of some recommended security measures.

Political Context

The draft emerges amid mounting U.S. criticism of European AI regulation. Recent comments from the Trump administration emphasized prioritizing AI development over safeguards, warning against regulatory overreach potentially stifling innovation.

EU policy shifts include shelving the AI Liability Directive and proposing regulatory simplification measures. However, with the AI Act implementation ongoing, balancing robust oversight with industry viability remains challenging.

Industry Perspectives

Prominent GPAI developer Mistral has publicly struggled with certain technical compliance requirements, working closely with regulators to find feasible solutions.

Regulatory Horizon

The European Commission's AI Office is concurrently developing supplementary guidance to clarify GPAI definitions and responsibilities. These forthcoming interpretations may significantly influence the practical implementation of the AI Act.