Back to Learning Area

Regulating Foundation Models

Whoever trains a foundation model suddenly becomes regulatorily visible in the EU.

  • Ethics
  • 12 min
  • Intermediate

Plain Language

This preview is in Plain Language. The whole portal is also available in Plain Language. You can change the language at the top. Or use the button below. Our Plain Language texts are AI-assisted and editorially reviewed. They follow the core principles of easy-to-read language, but they are not certified Easy Read: a review panel of people with learning difficulties does not test them.

Preview: AI rules in Europe

What is general-purpose AI?

Some AI programs can solve many tasks. AI means Artificial Intelligence. This is called general-purpose AI. General-purpose means: for many purposes. Examples are GPT-4 and Claude.

The EU AI Act

The EU has a law for AI. It is called the EU AI Act. The law divides AI into levels. Risky AI has stricter rules. Harmless AI has lighter rules.

Who is responsible?

The manufacturer builds the AI model. They are responsible for the base. The company uses the model. They are responsible for the use.

When the EU designed its AI Act around risk classes, it only gradually recognized the problem with general-purpose AI — in late 2022, ChatGPT made it obvious to everyone. A model like GPT-4, which followed a few months later and writes legal briefs AND generates memes AND assists in medical diagnosis, cannot be classified as "high risk" or "minimal risk" because it is both simultaneously.

The solution? Regulate the foundation, not the building. This article explains how Chapter V of the AI Act creates a dedicated rulebook for General Purpose AI providers — what they must disclose, when enhanced duties kick in, why open-source models get a deliberate advantage, and how the European Commission has been able to enforce these rules with fines since August 2026.

The Central Insight

GPAI models are chameleons — their risk depends entirely on how they are deployed. The AI Act responds by splitting responsibility: Providers (the makers of the model) ensure the base model is transparent and documented. Whoever builds the model into their own application (downstream providers) or uses that application (deployers, i.e. operators) ensures that the specific application meets the relevant risk-class requirements. This dual-responsibility architecture allows one set of rules to cover a technology whose uses nobody can know in advance.

The GPAI Problem — Why Risk Classes Break Down

General Purpose AI (GPAI) refers to AI models trained on vast amounts of data for a wide range of tasks — without being tied to a single application. GPT-5, Claude, Llama, and Gemini are classic GPAI models. This very universality creates a regulatory problem.

Definition

General Purpose AI (GPAI)

Analogy

Think of a GPAI model as a universal motor onto which you can fit a saw, a drill, or a sander. The motor itself is neither safe nor dangerous; that depends on the attachment and on what you use it for. The motor manufacturer (Provider) must ensure that the motor runs safely and that the manual is complete. Whoever builds a saw from it and sells it (downstream provider) is responsible for the safety of the saw.

The risk-based approach of the AI Act works per specific application — but a GPAI model has no single fixed application. If you use a language model as a chat tool, only transparency obligations apply to that application: the chatbot must disclose that you are talking to an AI. If the exact same model becomes the core component of medical diagnostic software, the application immediately enters the high-risk category. The solution: Chapter V additionally regulates at the level of the model provider, independent of the use case.

Provider (e.g. Meta)

Develops and trains the base GPAI model. Responsible for the foundation: model safety, technical documentation, Chapter V duties. Example: Meta publishes Llama 3.1 8B.

Downstream provider (e.g. Startup)

Integrates the GPAI model into its own application and offers it. Responsible for the specific use case, e.g. an HR screening bot in the high-risk category. Example: Startup builds HR bot with Llama.

Caution

Misconception: Only EU companies are affected

Any provider that makes their model available on the European market falls under Chapter V — regardless of where the company is headquartered. OpenAI (US), Meta (US), Google (US): All must fulfill GPAI obligations when their models are available in the EU.

The Four Pillars of Transparency

All GPAI providers must fulfill four core transparency obligations. They have applied since 2 August 2025. Since 2 August 2026, the European Commission can also punish violations with fines. Models that were already on the market before 2 August 2025 have a transition period until 2 August 2027.

Technical Documentation Description of model architecture, training methods, internal evaluations, and known limitations — to be kept available for the AI Office and national authorities on request.
Training Data Summary Public, sufficiently detailed summary of training data: sources, scope, and curation methods. The AI Office published a mandatory template for it in July 2025.
Downstream Information Technical details for developers integrating the model into their own products.
Copyright Compliance A policy for complying with EU copyright law, including respect for opt-out rights (reservations of rights) against text and data mining.

Imagine a pharmaceutical company selling a raw active ingredient (the model) to other companies who formulate it into specific drugs. The ingredient manufacturer must provide a safety data sheet (technical documentation), disclose what the ingredient is made from (training data summary), provide formulation guidelines to downstream drugmakers (downstream information), and respect the rights of those whose material it uses (copyright compliance). If the ingredient is exceptionally potent (systemic risk), additional clinical safety trials (red teaming) and reporting duties apply.

FLOPs Threshold

10 ²⁵

Cumulative training compute above which a model is presumed to pose systemic risks.

Effective Since

Aug.2025

GPAI obligations have applied since 2 August 2025; fines have been possible since 2 August 2026.

Standard GPAI vs. Systemic Risk

Models whose training used more than 10²⁵ FLOPs are classified as models with systemic risk — unless the provider convincingly demonstrates that its model nevertheless poses no such risks. Anyone who reaches the threshold, or foreseeably will, must notify the European Commission without delay and within two weeks at the latest. These models face additional duties:

Standard GPAI

Four transparency obligations: Technical documentation, training data summary, downstream information, copyright compliance. Example: Llama 3.1 8B (far below threshold).

Systemic Risk

All four transparency obligations PLUS: Model evaluations including red teaming (adversarial testing), assessment and mitigation of systemic risks, incident reporting for serious events, enhanced cybersecurity. Example: GPT-4 (estimated by Epoch AI at around 2 × 10²⁵ FLOPs).

Concrete example: A model like GPT-4, whose training compute Epoch AI estimates at around 2 × 10²⁵ FLOPs (OpenAI itself has not published a figure), falls under the presumption of systemic risk. Its provider must conduct red teaming, assess and mitigate risks, document evaluations, and report serious incidents. A model like Llama 3.1 8B, by contrast, trained far below the threshold, qualifies as standard GPAI with only four baseline duties.

Caution

Misconception: 10²⁵ FLOPs is a permanent threshold

The 10²⁵ FLOPs threshold is not an immutable law of nature — in two respects. First, the European Commission can adjust it by delegated act as technology evolves. Second, it only establishes a presumption: The Commission can also classify models below the threshold as models with systemic risk if they have comparable capabilities or impact — on its own initiative or after an alert from its scientific panel. The yardstick is the criteria in Annex XIII, such as the number of parameters, the size of the dataset, or the reach on the EU market.

Who checks compliance depends on whether the model and the application come from the same hand. GPAI models are overseen by the European Commission's AI Office. Applications are normally supervised by national authorities. But if a provider builds the application on its own GPAI model — for example OpenAI with ChatGPT running on its own GPT models — then, since the AI Omnibus (Regulation (EU) 2026/1744), the AI Office alone is competent, including for the application. Exceptions apply, among others, to AI in products such as medical devices or machinery. The AI Office is also solely competent for AI systems that are, or are part of, a very large online platform or search engine. This way one body supervises the model and the application together, instead of splitting supervision between the EU level and national authorities.

Open Source — Privilege with Limits

Open-source GPAI models deliberately receive reduced obligations. The legislator's reasoning: Whoever discloses weights and architecture is already largely transparent, and open models drive research and innovation.

Note

The Open Source Privilege

Providers who publish their model under a free and open-source license, make its weights, architecture, and usage information publicly available, and do not make money from the model itself benefit from significantly reduced requirements. If, for example, commercial use requires a license fee or access is only available for payment, the privilege no longer applies, according to the European Commission's guidelines.

  • Waived: Technical documentation for the AI Office and authorities
  • Waived: Downstream provider information

Still required:

  • Copyright policy and respect for opt-out rights
  • Publication of training data summary

Absolute ceiling: The privilege does NOT apply to models with systemic risk. Every such model — as a rule, every model above 10²⁵ FLOPs — must fulfill ALL obligations, regardless of its license.

Think of vehicle registration: For a production car, the manufacturer must submit a thick folder of technical documents to the authority. If a hobbyist publishes the complete blueprints of their car online so that anyone can check them, the authority does not ask for that folder again — the information is already out in the open. For a 1,000-horsepower supercar (systemic risk), that is not enough: crash tests and type approval are mandatory, no matter how open the blueprints are. And anyone who hands out their "open" blueprints only for a fee counts as a regular manufacturer again.

Meta publishes Llama 3.1 8B with open weights. The training compute is below 10²⁵ FLOPs, so the reduced obligations would be an option. Whether Llama actually gets the privilege is doubtful, though: its license restricts use, very large services need extra permission from Meta, and according to the Commission's guidelines that speaks against a free license. For the openly released Llama 3.1 405B at around 3.8 × 10²⁵ FLOPs the question does not even arise: above the threshold for systemic risk the privilege vanishes and full compliance is required, whatever the license.

Who Is Responsible for What?

Who is responsible when an open-source model is misused? The AI Act draws a clear line: Whoever builds and offers their own application on top of it is, as provider of that application, fully responsible for it — including the obligations of its risk class. The Provider (e.g. Meta) is responsible only for the base model. If a developer builds a harmful tool from Llama, the developer as provider of that application bears full responsibility. Incidentally, who is liable for damages under civil law is not governed by the AI Act but by general liability law, such as product liability law.

Caution

Misconception: The AI Act bans open source

The exact opposite is true: The AI Act explicitly privileges open source with reduced obligations. The limits: The privilege does not apply to models with systemic risk, nor to models the provider makes money from. And it only concerns the obligations of model providers — whoever builds a prohibited or high-risk application from an open model is subject to the full rules.

What does "sufficiently detailed summary" or "adequate cybersecurity" mean in concrete terms? This gap between legal text and practice is filled by the GPAI Code of Practice, a code of conduct drawn up by independent experts in a process organized by the EU AI Office. It was published on 10 July 2025; on 1 August 2025 the European Commission and the AI Board confirmed it as an adequate voluntary tool. It has three chapters: Transparency (with a Model Documentation Form) and Copyright for all GPAI providers, and Safety and Security only for providers of models with systemic risk. According to the Commission's list (as of October 2026), signatories include Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI, and OpenAI, while xAI signed only the Safety and Security chapter; Meta publicly declined to sign in July 2025. The Code is not an additional obligation: Anyone who adheres to it can use it to demonstrate that they meet their obligations. Anyone who does not sign must prove to the Commission by other means that they comply. Since 2 August 2026, this carries weight: The Commission can impose fines of up to 3% of worldwide annual turnover or 15 million euros, whichever is higher.

How large the disclosure gap is can be seen in Stanford University's Foundation Model Transparency Index. In its third edition from December 2025 — that is, after the GPAI obligations took effect — the assessed providers scored an average of only 40 out of 100 points. In 2024 the figure was 58; however, the researchers have since tightened their criteria and assessed partly different companies, so the values can only be roughly compared. IBM reached 95 points, xAI and Midjourney only 14. Providers reveal the least about training data and the compute used for training. One caveat: The index measures what is publicly available. Technical documentation under the AI Act, by contrast, goes to the AI Office and to downstream providers; what must be public is mainly the training data summary.

Why exactly 10²⁵ FLOPs? When the AI Act was being negotiated, this value captured only the most powerful models: According to Epoch AI, GPT-4 (March 2023) was the first model trained with more than 10²⁵ FLOPs. By June 2025, Epoch AI already counted over 30 such models from 12 developers. So a threshold for a single case has become a threshold for an entire class of models. Crucially, the European Commission has the explicit authority to adjust this threshold.

The problem cuts both ways. On the one hand, training efficiency keeps rising thanks to better algorithms and hardware: Future models could reach the same capabilities with fewer FLOPs and slip under the threshold — an argument for a lower threshold. On the other hand, more and more providers train above 10²⁵ FLOPs, so the presumption of systemic risk also hits models that are no longer anywhere near the frontier — an argument for a higher one. That is why the Commission can also draw on criteria other than the number, such as capabilities and reach.

An additional challenge: With distributed training across many data centers, precisely measuring cumulative FLOPs becomes increasingly difficult. The methodology for compute measurement is itself the subject of scientific discussion.

Interactive: From the Data Protection Directive to the AI Act

Step through more than 30 years of European regulatory history — from the first Data Protection Directive in 1995 to the enforcement of GPAI obligations from 2026. Each milestone shows how the EU gradually built the legal framework for AI.

Click on a point in time to learn more.
1995

EU Data Protection Directive

Directive 95/46/EC was the first EU-wide legal framework for protecting personal data. It established core principles like purpose limitation, data minimization, and the right of access that continue to shape European data protection law today.

Significance: Created the foundation for the European data protection model, which was later evolved into the GDPR.

Note

Key Takeaways

  1. GPAI models escape standard risk classification because the same model can serve minimal-risk and high-risk applications — Chapter V solves this by regulating at the provider level.
  2. All GPAI providers must fulfill four transparency obligations (technical documentation, training data summary, downstream information, copyright compliance); models with systemic risk face additional duties (red teaming, risk assessment and mitigation, incident reporting, cybersecurity).
  3. Open-source GPAI enjoys reduced obligations — but this privilege vanishes immediately once the model counts as a model with systemic risk.
  4. The GPAI obligations have applied since 2 August 2025, and since 2 August 2026 the Commission can enforce them with fines; models already on the market before then have until 2 August 2027. The voluntary Code of Practice is a recognized way to demonstrate compliance.

Check Your Understanding

Can you explain this in your own words? Tick what you've got.

0 of 3 done

Check

Quiz: Foundation Model Regulation

What does "General Purpose AI" (GPAI) mean under the EU AI Act?

Select one answer

Answer Key: 1) B · 2) C · 3) B · 4) C · 5) C · 6) B