General-Purpose AI & Foundation Models

The risk tiers were written for AI systems with a fixed purpose. Foundation models broke that mould - one model, countless uses - so the Act added a separate regime for general-purpose AI. If you train, fine-tune, or distribute a foundation model, this lesson is your obligation list.

Why GPAI Needed Its Own Rules

The risk tiers classify a system by what it is for. A foundation model has no single purpose - the same large language model can write marketing copy, triage support tickets, or assist a medical workflow. Regulating it by use-case is impossible at the model layer, because the model provider does not control how downstream developers will deploy it. So the Act created a parallel track: obligations that attach to the general-purpose AI model itself, regardless of where it ends up. These rules became applicable in August 2025.

The Baseline: Obligations for All GPAI Providers

Every provider of a general-purpose AI model placed on the EU market must:

1. Maintain technical documentation of the model - its training and testing process and evaluation results - available to the AI Office on request.
2. Provide information to downstream providers who integrate the model, so they can understand its capabilities and limitations and meet their own obligations.
3. Put in place a copyright policy that respects EU copyright law, including honouring the text-and-data-mining rights reservations (opt-outs) that rightsholders have expressed.
4. Publish a sufficiently detailed summary of the content used to train the model, following the template provided by the AI Office.

Providers of models released under a genuinely free and open-source licence get a partial exemption from the documentation and downstream-information duties - but not from the copyright policy and training-data summary, and not at all if the model has systemic risk.

The Heavier Track: GPAI with Systemic Risk

A small number of the most capable models are designated as having systemic risk - the kind of model whose failures or misuse could ripple across the economy or society. A model is presumed to have systemic risk when the cumulative compute used to train it crosses a very high threshold (set in the Act at 10²⁵ floating-point operations), and the AI Office can designate others on a case-by-case basis.

Providers of systemic-risk models carry the baseline duties plus:

• Model evaluation, including standardised and adversarial testing (red-teaming) to identify and mitigate systemic risks;
• Systemic-risk assessment and mitigation across the model lifecycle;
• Serious-incident tracking and reporting to the AI Office and relevant authorities;
• Adequate cybersecurity for the model and its physical infrastructure.

Codes of Practice

Because the GPAI rules are new and technical, the Act leans on codes of practice - drawn up with the AI Office and industry - as the practical bridge to compliance until harmonised standards exist. Adhering to an approved code is a way for providers to demonstrate compliance with the GPAI obligations. If you provide a model, following the relevant code of practice is currently the most concrete path to showing you have met your duties.

What to Do If You Touch a Foundation Model

1. Determine your role. API caller (downstream)? Fine-tuner (possibly a provider)? Original trainer (provider)?
2. If you are a GPAI provider, build the documentation, downstream-information pack, copyright policy, and training-data summary now - these have applied since August 2025.
3. Check the systemic-risk threshold. If your training compute approaches the threshold, the heavier track applies and you should engage with the AI Office early.
4. If you are downstream, capture the information the model provider supplies - you will need it to classify and document your own end system under the tiers.