GPAI Models vs Foundation Models: EU AI Act Terminology Explained

One Technology, Many Names

Large models trained on broad data and adaptable to many tasks have collected an unusual number of names: foundation models, frontier models, general-purpose AI, dual-use foundation models, large language models. These labels are not interchangeable in regulatory documents — each is a defined term in a specific legal instrument, with obligations hanging off the definition. For anyone doing compliance work across jurisdictions, terminology is not pedantry; it determines which rules apply to which artefacts. This article maps the vocabulary, starting with the term that matters most in Europe.

What the EU AI Act Actually Says

Regulation (EU) 2024/1689 regulates the general-purpose AI model, defined in Article 3(63) as an AI model — including where trained with a large amount of data using self-supervision at scale — that displays significant generality, is capable of competently performing a wide range of distinct tasks regardless of how it is placed on the market, and can be integrated into a variety of downstream systems or applications, excluding models used for research, development or prototyping before being placed on the market. Two related terms complete the family. Article 3(66) defines the general-purpose AI system: an AI system based on a GPAI model that can serve a variety of purposes, directly or integrated into other systems. And Article 3(65) defines systemic risk, creating the subcategory of GPAI models with systemic risk under Article 51 — presumed where cumulative training compute exceeds 10^25 FLOPs, or following Commission designation.

Nowhere in the operative text does the adopted regulation regulate a foundation model. That absence is the residue of a legislative argument worth understanding.

The Legislative History: Foundation Model Rises and Falls

The term foundation model entered the debate from research: a 2021 Stanford report popularised it for models trained on broad data at scale and adaptable to a wide range of downstream tasks. When the European Parliament adopted its negotiating position in June 2023, it grafted the concept onto the draft regulation, proposing a dedicated regime for foundation models with obligations on risk mitigation, data governance and downstream documentation, plus extra duties for generative systems.

In the trilogue negotiations of late 2023, the framing changed. Co-legislators settled on general-purpose AI model — language closer to the Council's approach — with a two-tier structure: baseline transparency duties for all GPAI models, and systemic-risk duties for the most impactful ones, anchored to the compute presumption. The choice was substantive, not stylistic. Foundation model describes how a model is built and positioned in a technology stack; general-purpose describes what it can do. The adopted definition selects models by capability breadth — significant generality, wide range of distinct tasks — making the regime more robust to architectural change: whatever replaces today's transformer-based pre-training, a model with broad competence still meets Article 3(63).

The American Vocabulary: Dual-Use Foundation Models

United States usage went the other way. The October 2023 Executive Order 14110 on safe, secure and trustworthy AI defined the dual-use foundation model: an AI model trained on broad data, generally using self-supervision, containing at least tens of billions of parameters, applicable across a range of contexts, and exhibiting — or being modifiable to exhibit — high performance at tasks posing serious security risks such as lowering barriers to chemical, biological, radiological or nuclear weapons development or advanced cyber operations. Its reporting triggers used compute thresholds an order of magnitude above the EU presumption — 10^26 FLOPs for general models. The subsequent change of administration in 2025 rescinded and reworked parts of that framework, which is itself a terminology lesson: US obligations have shifted with executive instruments, while the EU's definitions sit in a regulation with the permanence of statute.

The Safety Community's Word: Frontier Models

Frontier model — favoured in the UK's AI Safety Summit process, company safety frameworks and research discourse — denotes the most capable models at the leading edge of development. It is a moving target by design: today's frontier is next year's commodity. The UK has used it in voluntary commitments rather than binding statute, and the term has no operative role in the EU regulation. Its nearest EU relative is the GPAI model with systemic risk, but the mapping is imperfect: the EU category is legally fixed by compute presumption and designation, while frontier status is a relative, informal judgement. A model can leave the frontier in capability terms yet remain on the Commission's systemic-risk list.

Why the Differences Matter in Practice

Four practical consequences flow from the vocabulary map:

• Scope mismatches: a model can be a GPAI model in the EU while falling outside US reporting thresholds, and vice versa. Compliance matrices must be built per instrument, not per buzzword.
• Threshold confusion: the EU uses 10^23 FLOPs as an indicative criterion for being a GPAI model at all (in the Commission's 2025 guidelines) and 10^25 FLOPs for the systemic-risk presumption; the 2023 US order used 10^26. Teams that quote a single FLOPs number without naming the instrument routinely brief their leadership wrongly.
• Contract drafting: agreements that promise compliance for foundation models leave open whether the EU regime is meant. Precise contracts reference the GPAI model definition of Article 3(63) and the systemic-risk classification of Article 51.
• Communication discipline: marketing copy calling a product a frontier foundation model invites regulators to test whether the systemic-risk criteria of Annex XIII — capabilities, benchmarks, reach — are met. Public capability claims are evidence.

Model Versus System: The Second Distinction

The EU vocabulary embeds a further split that other frameworks blur: the model is not the system. Chapter V obligations attach to the GPAI model and its provider; everything else in the regulation — transparency under Article 50, the high-risk regime, deployer duties — attaches to AI systems. A chat product is a GPAI system built on a GPAI model; one company may be provider of both, two companies may split the roles. Borrowing American or research vocabulary that treats the model and the product as one artefact obscures exactly the boundary on which EU obligations divide.

Other Jurisdictions, Briefly

The vocabulary fragmentation extends beyond the three main frameworks. Korea's AI Basic Act works with a concept of high-impact AI and addresses generative AI explicitly, defining categories that overlap with but do not mirror the EU's. China regulates by function rather than model class — its generative AI measures and deep synthesis rules attach to services, with algorithm and model filing duties. Japan's AI promotion framework avoids binding model-class definitions altogether, favouring sectoral guidance. International soft-law processes — the G7 Hiroshima process and the OECD — generally say advanced AI systems precisely to avoid choosing among national vocabularies. For multinationals, the operational consequence is that a single internal model register must carry several classification columns, one per jurisdiction, each anchored to the local statutory text rather than to a shared informal label.

A Concrete Example

A compliance lead at a multinational is asked whether the company's new in-house model is a foundation model and what that means legally. The precise answer runs: under the EU AI Act it is a GPAI model, because its training compute exceeds the indicative 10^23 FLOPs criterion and it generates text across a wide task range; it is far below 10^25 FLOPs, so no systemic-risk presumption arises; the company places it on the EU market inside its products, so Article 53 duties attach to the group entity doing so. Under the current US framework, no compute-based federal reporting trigger is met. In the company's model risk documentation, frontier model is avoided entirely; the internal register records the Article 3(63) analysis and the compute figure. One question, three vocabularies, one defensible answer.

Action Plan

Standardise vocabulary internally: use GPAI model and GPAI system when discussing EU obligations, reserve foundation model for technical architecture discussions, and treat frontier model as a research label with no compliance content. Maintain a glossary in your AI governance policy mapping each term to its instrument and threshold, and train legal, engineering and communications teams on it. Terminology discipline costs nothing and prevents the two most expensive errors in this field: assuming a rule does not apply because the marketing name differs, and publicly claiming capabilities that put you in a category your compliance programme has not prepared for. When in doubt, quote the statute: the definition in Article 3(63) is forty words long, and citing it verbatim settles more internal debates than any amount of paraphrase. The glossary habit also future-proofs training materials: definitions quoted from statute survive regulatory amendments far better than paraphrases written around last year's product names.

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.