EU AI Act Enforcement Kicks Off With Foundation Model Rules

The European Union has officially begun enforcing key provisions of its landmark AI Act, marking the start of the world's most comprehensive regulatory framework for artificial intelligence. Foundation model providers — including companies like OpenAI, Google, Meta, and Anthropic — now face strict transparency, documentation, and risk-management requirements that could reshape how AI is developed and deployed globally.

The enforcement phase, which kicked in following a phased rollout timeline established when the AI Act was formally adopted in 2024, targets general-purpose AI (GPAI) models with particular scrutiny reserved for systems deemed to pose 'systemic risk.' Companies that fail to comply face fines of up to €35 million or 7% of global annual revenue — whichever is higher.

Key Facts at a Glance

• Foundation model providers must now publish detailed technical documentation, including training data summaries and energy consumption reports
• Models classified as posing systemic risk (trained with compute exceeding 10^25 FLOPs) face additional obligations including adversarial testing and incident reporting
• Fines can reach up to €35 million or 7% of global revenue for non-compliance
• The newly established EU AI Office in Brussels serves as the primary enforcement body for GPAI rules
• Copyright holders gain new rights to information about whether their works were used in training data
• All providers must implement policies to comply with EU copyright law, regardless of where they are headquartered

Foundation Models Face Unprecedented Scrutiny

The AI Act divides general-purpose AI models into 2 tiers. Standard GPAI models must meet baseline transparency requirements. These include publishing sufficiently detailed summaries of training data, maintaining up-to-date technical documentation, and implementing policies that respect EU copyright frameworks.

The second tier — systemic risk models — triggers a far more demanding compliance regime. Any model trained using computational power exceeding 10^25 floating-point operations (FLOPs) automatically falls into this category. Currently, that threshold captures frontier models such as OpenAI's GPT-4, Google's Gemini Ultra, and potentially Anthropic's Claude 3.5 Opus.

Systemic risk providers must conduct and document model evaluations, including red-teaming exercises and adversarial testing. They are also required to track and report serious incidents, implement cybersecurity protections for model weights, and report energy consumption metrics. Unlike the self-regulatory approaches favored in the United States, these are legally binding obligations backed by significant financial penalties.

The EU AI Office Takes Center Stage

Enforcement of GPAI-specific rules falls to the EU AI Office, a dedicated body established within the European Commission in Brussels. The office, which began operations in early 2024 with approximately 140 staff members, serves as the central hub for monitoring foundation model compliance across all 27 EU member states.

The AI Office has the power to request documentation, conduct evaluations, and ultimately levy fines against non-compliant providers. It also coordinates with national authorities, who handle enforcement of other AI Act provisions related to high-risk AI systems in areas like healthcare, law enforcement, and education.

Industry observers note that the office's relatively small staff could face challenges in policing the rapidly expanding global AI market. However, the Commission has signaled plans to expand the team and has established advisory bodies — including a scientific panel of independent experts — to support its technical assessment capabilities.

How US AI Giants Are Responding

American tech companies, which dominate the foundation model landscape, have been preparing for this moment for months. OpenAI published a preliminary EU compliance report earlier this year, outlining its approach to training data transparency and safety evaluations. Google DeepMind has similarly expanded its model documentation practices.

Meta faces a particularly complex situation. Its open-source Llama models are freely available for download, raising questions about how open-weight releases interact with the AI Act's documentation and copyright compliance requirements. The regulation does include some accommodations for open-source models, exempting them from certain obligations — but not if they are classified as posing systemic risk.

Key compliance challenges for US providers include:

• Providing training data summaries that satisfy EU requirements without revealing proprietary information
• Meeting copyright compliance obligations in a legal landscape that differs significantly from US fair-use doctrine
• Establishing EU-based points of contact and representation for regulatory communications
• Implementing incident reporting pipelines that meet the AI Office's timeline requirements
• Conducting energy consumption assessments for models trained on massive GPU clusters
• Navigating potential conflicts between EU transparency rules and trade secret protections

Anthropic, maker of the Claude model family, has taken a somewhat different approach by emphasizing its existing Responsible Scaling Policy as a framework that aligns with EU expectations. However, whether voluntary corporate policies satisfy legally binding regulatory requirements remains an open question that the AI Office will likely need to address.

Comparing EU and US Regulatory Approaches

The enforcement of the AI Act throws the stark divergence between European and American regulatory philosophies into sharp relief. While the EU has opted for a comprehensive, legally binding framework, the United States continues to rely primarily on executive orders and voluntary industry commitments.

President Biden's 2023 Executive Order on AI Safety established reporting requirements for frontier models, but these lack the statutory force and penalty structures of the EU approach. The current US administration has shown even less appetite for binding AI regulation, preferring to let market dynamics and industry self-governance drive safety practices.

This regulatory gap creates a complex compliance landscape for multinational AI companies. Models deployed in the EU must meet the AI Act's requirements, while the same models face far fewer obligations in the US market. Some industry analysts worry this could lead to a 'regulatory arbitrage' dynamic, where companies invest less in safety and transparency for markets without binding rules.

Conversely, others argue the EU's approach risks stifling innovation. The compliance costs associated with detailed documentation, adversarial testing, and energy reporting could disadvantage smaller European AI startups compared to well-resourced American incumbents. A recent estimate from the Centre for European Policy Studies suggested that full AI Act compliance could cost large providers between $5 million and $15 million annually, with smaller firms facing proportionally heavier burdens.

What This Means for Developers and Businesses

For AI developers building on top of foundation models, the AI Act's enforcement creates both obligations and opportunities. Downstream deployers — companies using APIs from OpenAI, Google, or Anthropic to build products — inherit certain transparency requirements that must be passed through the value chain.

Practical implications include:

• AI-generated content must be labeled as such in many contexts, requiring developers to implement watermarking or disclosure mechanisms
• Companies deploying AI in high-risk domains (hiring, credit scoring, education) must conduct conformity assessments and register systems in the EU database
• Chatbot providers must ensure users know they are interacting with AI, not a human
• Businesses using AI for emotion recognition or biometric categorization face outright bans in many contexts

For European startups, the regulation creates a potential competitive advantage in 'compliance-ready' AI products. Companies that build transparency and documentation into their development processes from day one can market themselves as AI Act-compliant — a selling point that may carry increasing weight as enforcement ramps up.

Looking Ahead: Phased Enforcement Timeline

The foundation model rules represent just one phase of the AI Act's staggered enforcement timeline. The full rollout continues through 2026, with additional provisions taking effect at regular intervals.

Prohibited AI practices — including social scoring systems, untargeted facial recognition scraping, and manipulative AI techniques — were among the first provisions to take effect. High-risk AI system requirements, covering sectors like healthcare, transportation, and law enforcement, are scheduled for enforcement by August 2026.

The AI Office is expected to publish its first codes of practice for GPAI models in the coming months. These codes will provide more granular guidance on how providers can demonstrate compliance with the Act's general obligations. Companies that adhere to approved codes of practice will benefit from a 'presumption of conformity' — essentially a safe harbor that simplifies regulatory interactions.

Industry watchers should also monitor the AI Office's approach to the systemic risk threshold. The current 10^25 FLOPs benchmark was acknowledged as a blunt instrument, and the Commission has the authority to update it as the technology evolves. With training compute costs dropping rapidly — driven by more efficient architectures and hardware — the number of models crossing the systemic risk threshold is likely to grow substantially over the next 18 months.

The EU AI Act represents the most ambitious attempt by any government to regulate artificial intelligence comprehensively. Its enforcement sends a clear signal: the era of unregulated AI development is ending, at least in Europe. Whether this framework becomes a global template — as GDPR did for data privacy — or an isolated regulatory experiment will depend on how effectively the AI Office balances innovation incentives with genuine safety oversight in the months and years ahead.