EU AI Act Enforcement Begins: Fines Up to €15 Million for Foundation Model Providers

The European Commission gained enforcement authority over general-purpose AI models on August 2, 2026. This marked the shift from a collaborative grace period to binding enforcement backed by fines and regulatory action. The obligations are detailed in a Code of Practice, published in July 2025, that establishes accountability frameworks for foundation model developers—the companies creating adaptable systems like ChatGPT and Claude that can be deployed across multiple downstream applications. Over 1,000 stakeholders, including model providers, civil society organizations, academics and trade associations, participated in drafting the Code through four working groups over about nine months.

For US AI companies doing business in Europe, the enforcement timeline matters now. The Commission can demand access to models, request technical documentation, impose corrective measures, and fine noncompliant providers up to €15 million or 3% of global turnover. Models released before August 2, 2025 get a two-year transition until August 2, 2027 to achieve full compliance, but the Commission expects concrete progress. A grace period that encouraged signatories to implement measures voluntarily through August 1, 2026 has closed, and enforcement began August 2.

What the Code of Practice is and how it works

The Code of Practice is a voluntary compliance framework designed to bridge the gap between when general-purpose AI obligations took effect on August 2, 2025 and when formal EU standards are expected in 2027 or later. While technically not legally binding, providers adhering to the Code can demonstrate compliance with their obligations under Articles 53 and 55 of the EU AI Act. Non-signatories must prove compliance through other means, potentially facing more rigorous Commission scrutiny. The Commission has stated that for providers of general-purpose AI models adhering to an assessed adequate code of practice, it will focus its enforcement activities on monitoring adherence to the code rather than investigating from scratch.

The Code was developed through an extensive multi-stakeholder process beginning in October 2024, with thirteen independent Chairs and Vice-Chairs leading four working groups. The drafting process included three public consultation rounds before the final version was published July 10, 2025. The Code emphasizes accountability, transparency, and risk management proportionate to model capabilities—not treating all general-purpose models as equally dangerous.

Who counts as a general-purpose AI provider

The EU AI Act defines general-purpose AI models as adaptable, pre-trained systems that can be applied across multiple downstream uses. The defining trait is versatility: the same model serves as a customer-service chatbot one day and a coding assistant the next. This contrasts with purpose-built systems trained for a single specific task. Foundation model developers—companies creating these base models—are the primary targets.

A critical exemption exists for open-source models released under truly free licenses from some documentation and representation requirements. However, this exemption does not apply to models classified as having systemic risk—high-impact systems that the Commission identifies as presenting serious risks to fundamental rights or public safety. Those models must comply with full requirements regardless of licensing. Systemic risk refers to models with broad capabilities that could cause widespread harm if misused or malfunction, affecting many people across multiple applications. The Commission also operates an AI Act Service Desk that provides information and support to help organizations implement the AI Act.

Transparency requirements: documentation and public disclosure

Providers must maintain comprehensive technical documentation for every model using a standardized Model Documentation Form that covers architecture, training methodologies, data provenance, computational resources, and energy consumption. Documentation must be retained for at least ten years and provided to downstream users and authorities upon request. This requirement exists because downstream developers, regulators and users need to understand the model's foundational knowledge, training process, and potential biases. The documentation requirements mean companies must track not just what data was used, but how that data was obtained, processed, and selected for training.

The Code of Practice requires providers to publish a summary of the content used for training their models—effectively disclosing which data sources informed the system and how much data came from each source. Providers must also specify intended and unintended use cases, helping downstream developers understand where the model will work reliably and where it may fail.

Copyright compliance and lawful data sourcing

Developers must establish copyright policies with clearly defined internal responsibilities, ensuring they used only lawfully available content during model training. In practice, this means respecting machine-readable rights signals like robots.txt files that creators use to indicate their content should not be scraped, and honoring legal licenses and restrictions. Providers must implement technical protective measures designed to minimize the likelihood that models produce copyright-infringing content. The Code does not mandate which technologies, but it requires active, systematic approaches using state-of-the-art rigor—not merely hoping the model happens to respect copyright by accident.

Copyright requirements apply universally, eliminating any license exemption for open-source providers. Companies must designate contact points where copyright holders can report concerns and have processes to address complaints promptly. This creates ongoing accountability: if a model produces content matching copyrighted material, holders can notify the provider and request remediation.

Safety assessment and incident reporting for advanced models

For models posing systemic risk—those with high impact on rights, safety or security—developers must conduct risk assessment before market release, identifying potential harms, analyzing their severity and likelihood, and determining whether residual risks are acceptable. Providers must prepare detailed Safety and Security Model Reports describing testing methodologies, findings and mitigation measures. These reports must document how the model was evaluated for risk, including the analysis and mitigation measures applied. Providers must implement ongoing monitoring including filtering, monitoring systems, and refusal training where the model learns to decline problematic requests. This continuous cycle means the responsibility does not end at release.

Serious incidents must be reported to authorities within 2 to 15 days depending on severity. Governance requirements mandate that companies clearly assign oversight, ownership and assurance responsibilities within their organizations, allocate adequate resources, and protect whistleblowers reporting safety concerns internally. Unlike stricter regulations that ban certain AI uses entirely, the Code allows continued operation if providers systematically manage risk through documentation, monitoring, testing, and transparent reporting. The approach assumes that general-purpose models will be deployed widely, so the goal is managing their risks rather than preventing their use.

The grace period: what changed August 2, 2026

From August 2, 2025 to August 1, 2026, the AI Office adopted a collaborative approach with Code signatories. The Code stated that if providers did not fully implement all commitments immediately upon signing, the AI Office would not view these providers as having breached their commitments. This grace period allowed companies time to build the systems, hire the staff, and establish the processes needed for compliance. Providers could sign the Code, begin implementation, and work with regulators to identify practical obstacles without facing enforcement action.

Starting August 2, 2026, that flexibility ended. The EU Commission enforces all obligations for providers and may impose fines for any noncompliance—up to €15 million or 3% of global turnover, or €35 million or 7% of turnover for prohibited practices. The shift is not from voluntary to mandatory; the Code was always a demonstration of compliance with mandatory obligations. Rather, enforcement shifted from monitoring and collaboration to investigation and punishment. This means companies that have not yet signed the Code, or that signed but made minimal progress, face immediate regulatory scrutiny. Noncompliant providers may receive orders to correct violations, demands to halt certain model operations, or fines. The transition period for models already on the market extends to August 2, 2027, but that is not a reprieve from enforcement—it is a deadline by which full compliance must be achieved.

Enforcement powers and financial penalties

The AI Office and national authorities now hold investigative powers to request information from providers, demand model access for evaluation, require corrective measures, and issue fines. They can ask for any technical documentation, testing results, incident reports, or training data summaries providers maintain. They can evaluate models directly to assess compliance with transparency, copyright and safety requirements. If they find violations, they can order providers to cease certain practices, modify model behavior, or implement new safeguards. These powers are broad because general-purpose AI models can affect large populations, and regulators need tools to investigate problems.

Financial penalties create strong compliance incentives. Administrative fines reach €15 million or 3% of annual global turnover—whichever is larger. For a company with $10 billion in annual revenue, 3% means a potential fine of $300 million. For prohibited practices—such as banned uses of AI like manipulative techniques, social scoring, or generating child sexual abuse material—fines climb to €35 million or 7% of turnover, potentially reaching $700 million. These penalties dwarf the costs of compliance, making proactive adherence far cheaper than noncompliance and litigation.

Related coverage: What Compliance Requirements Does New York's RAISE Act Impose on AI Model Developers?; What the RAISE Act Requires From Company Boards.