Is a GPAI model the same as a high-risk system?

No. GPAI is a separate track with its own obligations, which attach to the model rather than to a particular use. A downstream application built on a GPAI model can nonetheless be high-risk in its own right — for example an Annex III recruitment or credit-scoring system — in which case both sets of duties are in play at once, for different parties.

We fine-tune an existing model. Are we a GPAI provider?

Possibly. Placing a model on the market under your own name, or modifying it substantially, can put you in the provider position for that model. If you simply consume a model through an API without changing it, you are normally a deployer of the resulting system rather than a provider of the model — but you still need the upstream provider's documentation to satisfy your own obligations.

What does the systemic-risk tier add?

For GPAI models with systemic risk, the baseline duties are joined by model evaluation, adversarial testing, incident reporting and cybersecurity obligations. The reasoning is that a model capable of affecting many downstream systems at once has to be probed for its failure modes before those failures propagate.

Who supervises GPAI providers?

The AI Office at Union level supervises general-purpose AI models, while national market surveillance authorities handle AI systems more broadly. For a company building on top of a model, that means two different addresses: the upstream model sits with the AI Office, your own system sits with your national authority.

AI Act

General-purpose AI models under the AI Act

Q: What counts as a general-purpose AI model?

A model that displays significant generality and can competently perform a wide range of distinct tasks, and that can be integrated into a variety of downstream systems or applications. Large language models and large multimodal models are the obvious examples. The test is generality and reusability, not the underlying architecture or the name of the vendor.

The risk-based rules ask what an AI system is used for. That question does not work for a model that can be used for almost anything, so the AI Act adds a second track: obligations that attach to general-purpose AI models themselves, and heavier duties for the small set of models judged to carry systemic risk.

GPAI in 60 seconds

What it covers: General-purpose AI models — significant generality, a wide range of tasks, and integration into many downstream systems.
A separate track: GPAI duties attach to the model itself, alongside — not instead of — the risk-based rules that apply to AI systems.
Baseline duties: Technical documentation, information to downstream providers, a copyright policy, and a public summary of the training data.
Systemic risk tier: Additional duties: model evaluation, adversarial testing, incident reporting and cybersecurity.

What counts as a GPAI model

A general-purpose AI model displays significant generality, can competently perform a wide range of distinct tasks, and can be integrated into a variety of downstream systems or applications. Large language models and large multimodal models are the clearest cases.

The classification follows capability and reusability, not vendor size or architecture — a separate track from the risk-based approach that governs AI systems. A narrow model trained for one task is not GPAI, however large it is; a broadly capable model is GPAI even when it is used, today, for only one thing.

Baseline duties for every GPAI provider

These apply to providers of general-purpose AI models placed on the EU market.
Duty	What it involves
Technical documentation	Documentation of the model — its training and testing process, and the results of its evaluation — kept up to date and available to the authorities on request.
Information to downstream providers	Documentation that lets companies integrating the model understand its capabilities and limitations well enough to meet their own obligations.
Copyright policy	A policy to comply with Union copyright law, including respecting reservations of rights in text and data mining.
Training-data summary	A sufficiently detailed public summary of the content used to train the model.

The systemic-risk tier

A subset of GPAI models is treated as carrying systemic risk, on the reasoning that failures in a model many organisations build on propagate widely. Those providers keep the baseline duties and add:

Model evaluation against standardised protocols, to establish what the model can and cannot do.
Adversarial testing to find and document the ways the model can be pushed into harmful behaviour.
Incident reporting — tracking serious incidents and reporting them, with corrective measures, to the AI Office and where relevant national authorities.
Cybersecurity protection for the model and its physical infrastructure.

If you build on top of a GPAI model

Downstream reality check

Most organisations are not GPAI providers — they are downstream builders consuming a model through an API. Three consequences follow.

First, you depend on the upstream provider's documentation: the information they must give you is what lets you describe your own system accurately. Ask for it and store it with your file rather than assuming it will be available when an authority asks.

Second, your application is classified on its own merits. Building a recruitment screen or a credit decision on top of a general model makes your system high-risk regardless of how the model itself is regulated as a high-risk system or not.

Third, the line between deployer and provider is easy to cross. Putting your own name on the model, or modifying it substantially, can move the model-level obligations onto you.

Who supervises what

The AI Office supervises general-purpose AI models at Union level, while national market surveillance authorities supervise AI systems in their own member state. For a company building on a third-party model, that means the model sits with one supervisor and your own system with another — and it is your system, documented in your name, that your national authority will ask about.

Frequently asked questions

Sources

EUR-Lex — Regulation (EU) 2024/1689 (Artificial Intelligence Act)

AI Act document packages are available now

The AI system inventory, risk classification workbook and documentation templates, in three tiers from 99 EUR. Pay once, download immediately.

See the packages