The first article explained what a foundation model is: a general-purpose model trained on large amounts of broad data, which is then adapted for many different tasks. It also traced where the term came from, why researchers argued about it, and how it moved from academic papers into law.
Once a term enters legislation, its precise meaning becomes much more important. Several different terms are now used to describe these models. Foundation model, large language model (LLM), generative AI and frontier model are often treated as if they mean the same thing, yet they do not. Each term describes a different feature of an AI model and therefore answers a different question.
| Term | The question it answers |
|---|---|
| Foundation model | What sits underneath, and what is built on top of it? |
| Large language model (LLM) | What type of information does it work with? |
| Generative AI | What does it produce? |
| Frontier model | How capable is it compared with other models? |
1. Foundation model: What sits underneath?
A foundation model is a general model that can be used as the starting point for many different applications (Bommasani et al. 2021). Think of it as the base layer of a pyramid. One model sits underneath, and developers build different products and services on top of it. The important features are the model's generality and its relationship to whatever is built on it.
2. Large language model: What does it work with?
A large language model works primarily with language or text. The term tells us what kind of information the model handles. It does not tell us how the model was trained, how powerful it is, whether it is a foundation model, or what it is ultimately used for. Many well-known foundation models are also LLMs, which is why the two get confused despite answering different questions.
3. Generative AI: What does it produce?
Generative AI describes what a model or system does. Specifically, it generates new content. That content might be text, images, audio, video or code. A model may also be multimodal, meaning it can work with or generate more than one type of content, such as text, images and audio. The term describes the output or capabilities of the model or system, not how it was built.
4. Frontier model: How capable is it?
When a model is described as a frontier model, it means that it sits at or near the leading edge of AI capability. Unlike the other three, this category moves. A model considered frontier today may not be in a few years, as more capable models are built. Frontier is therefore not a separate type of AI. It describes where a model sits relative to the rest. One law, discussed below, turns this informal idea into a precise legal test.
Why these categories are not interchangeable
At first sight, it is tempting to think of these terms as four circles, one inside another. They are not, and some real-world examples show why.
BERT. Released by Google in 2018, BERT is a foundation model (Devlin et al. 2018; Bommasani et al. 2021). It was built as a general base for other applications, and Google Search itself came to rely on it (Nayak 2019). It is also a language model, because it works with text. However, it is not generative AI in the usual sense, since it was designed to interpret and classify language rather than produce new content. And while it was highly capable on release, it would not count as frontier today. One model, four questions, four different answers.
DALL·E. Developed by OpenAI, DALL·E is a foundation model and an example of generative AI, because it generates images from text prompts (Ramesh et al. 2021). It is not usually described as an LLM, because it is built to produce images rather than text.
A modern AI assistant, such as ChatGPT. This can fit all four categories at once. It is a foundation-model-based system, because its underlying models can support many different applications and tasks. It is built on LLMs, because those underlying models work extensively with language. It is generative AI, because it produces new content in response to prompts. Lastly, it can involve a frontier model, where the underlying model meets the relevant threshold.
A traditional spam filter. This answers no to all four. It is not a foundation model, because it is designed to perform the specific task of identifying unwanted email. It is not a large language model: although it processes language, it is a classification system rather than a general-purpose language model. It is not generative AI, because it identifies and sorts messages rather than generating new content. Lastly, it is not a frontier model, because it does not require the enormous training compute associated with models that meet a frontier-model threshold.
The point is that these terms answer different questions. Describing a model as an LLM tells you that it works with language. It does not tell you how powerful it is, how widely it is used, or whether other important systems depend on it.
How are these terms used in law?
For legislators, the existence of four overlapping terms presents a question of legal scope, because a law must establish the category of models or systems to which its obligations apply. It cannot treat all four concepts as interchangeable, because each captures a different characteristic of a model.
Defining the relevant category therefore requires answering several questions. Should the law distinguish models by how they are developed, what they work with, what they produce, or the capabilities they possess? That choice has legal consequences.
A definition based on one characteristic may bring some models within the scope of the law while leaving others outside it. Whatever the definition does not capture may fall outside the relevant obligations. Three jurisdictions have confronted this definitional question, and each has taken a different approach.
1.The European Union: The generality question
The European Union built its rule around the question of what a model is and what can be built on it. The European Parliament used the term "foundation model" in its negotiating position of June 2023. The term was dropped during negotiations, and the AI Act as adopted regulates general-purpose AI models instead.
The definition in Article 3(63) focuses on generality. It asks whether a 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.
The Act then creates a second category for general-purpose AI models judged to carry systemic risk. Under Article 51, a model is presumed to fall into that category when the cumulative computing power used to train it exceeds 10²⁵ floating point operations. Crossing that line has immediate consequences. Under Article 52, the provider must notify the European Commission within two weeks. The model then carries a set of obligations in addition to those that apply to general-purpose AI models generally.
The EU is therefore asking two questions in sequence. Is this model general enough to be regulated at all, and if so, is it powerful enough to carry systemic risk?
2. California: The capability question, made precise
California built its rule around a different question. It asks how capable a model is, rather than what it is or what can be built on it, and so turns the informal idea of a frontier into a legal test. The Transparency in Frontier Artificial Intelligence Act, signed on 29 September 2025, defines a foundation model first. It then defines a frontier model as a foundation model trained using more than 10²⁶ integer or floating-point operations. The layering here is noteworthy. Frontier is not a separate kind of model but a label applied to a foundation model that has crossed a line, as the BERT example suggested.
Two details matter in the California law. First, the calculation includes computing power used for later fine-tuning, reinforcement learning and other material modifications, not just the original training run. Second, the California Department of Technology must assess technological developments, scientific literature and relevant standards each year. It must then recommend whether the definitions, including the threshold, should be updated to remain appropriate as the technology develops.
3. The United States federal government: Several questions at once (now revoked)
The federal government issued Executive Order 14110 on 30 October 2023. It defined a dual-use foundation model by reference to training on broad data, self-supervision, at least tens of billions of parameters, applicability across a wide range of contexts, and certain national-security capabilities. Although it was revoked in January 2025, it remains instructive. The definition asked three things at once: how the model was built, how large it is, and what it is capable of. The EU and California each chose one question; this definition combined them.
Comparing the EU and California compute thresholds
Both the EU and California use training compute, the amount of computing power used to build a model, to identify particularly powerful AI models. However, they use this measure in different ways.
The numbers are different. The EU uses 10²⁵ operations for its systemic-risk threshold. California uses 10²⁶ operations for its frontier-model threshold. California's threshold is therefore ten times higher.
The thresholds do different things. In California, crossing 10²⁶ operations is what makes a foundation model a frontier model, and that is what brings it within the relevant frontier AI obligations.
In the EU, crossing 10²⁵ operations does not determine whether a model is regulated as a general-purpose AI model. That is determined first by the model's generality under Article 3(63). The compute threshold then determines whether the model is presumed to pose systemic risk and therefore faces additional obligations.
The EU does use compute at the entry stage as well, but at a far lower level and only as guidance. The Commission's Guidelines treat a model as a general-purpose AI model where its training compute exceeds 10²³ operations and it can generate language (text or audio), or images or video from text. That figure is an indicative criterion in non-binding guidance, not a threshold in the Act itself.
The EU threshold is a presumption, while California's is a boundary. The EU's 10²⁵ figure creates a presumption of systemic risk. A provider can argue that its model, despite crossing the threshold, does not present systemic risk. The Commission can also designate a model as posing systemic risk even where it falls below the threshold. Training compute must be estimated rather than measured directly, and the Guidelines expect that estimate to be accurate to within 30 per cent.
California takes a more straightforward approach. Its compute threshold is a fixed boundary: a model is either above it or below it.
They treat later modifications differently. California is asking whether a model is powerful enough to qualify as a frontier model. It therefore counts computing power used for later fine-tuning and other material modifications towards the threshold, so a model could cross the threshold after its original training.
The EU is asking a different question: who is responsible for a modified model? Under its Guidelines, a model remains the same model when its original provider modifies it. However, if another company significantly modifies someone else's model, that company can become a provider in its own right. One indicative measure of a significant modification is computing power exceeding one third of the original model's training compute.
California adds another filter. One further difference has nothing to do with the model's capabilities. Many of California's obligations apply only to large frontier developers, meaning companies that, together with their affiliates, had more than $500 million in gross revenue in the preceding calendar year. A company could therefore develop a model that exceeds California's compute threshold and still fall outside many of the law's obligations, because the company itself is not large enough.
The key similarity is that both jurisdictions use compute as a proxy for capability. The EU uses it to identify systemic risk within a category that is already regulated. California uses it to draw the boundary of the category itself.
Why this matters
A model trained using 5 × 10²⁵ operations falls on different sides of the two lines. It is above the EU systemic-risk threshold, so it is presumed to have high-impact capabilities and attracts additional obligations. It is below the California threshold, so it is not a frontier model under that law at all.
The model has not changed. The legal result changes because the two laws draw the line in different places, for different purposes.
For African policymakers, three practical points follow. They matter most where a jurisdiction is borrowing a definition rather than writing its own.
First, every definition focuses on something different, so adopting a term means adopting what it measures. "Frontier model" focuses on scale. "General-purpose AI model" focuses on generality. "Foundation model" focuses on the model's role as a base for others, which carries its own concern, since a fault in the model underneath can spread to everything built on top of it.
Second, numerical thresholds go out of date, and the jurisdictions that set them acknowledge this. Both the EU and California built in ways to change their figures. The European Commission can amend its thresholds through delegated acts, and California requires an annual review of its definitions. The EU mechanism is already in use. In August 2025, the Commission stated that its 10²⁵ systemic-risk threshold was under review (European Commission 2025). A jurisdiction that copies a number without also creating a way to change it will end up with a rule it cannot fix.
Third, these definitions are proxies for harm. Compute, parameter count and generality are not harms. They are measurable stand-ins used to identify models that may deserve closer attention. That makes them practical but imperfect. A system can cause real harm in a specific application without crossing any threshold.
So it is not enough to look at the number in a legal definition. The questions to ask include where the line is, what counts towards it, what happens when a model crosses it, and what the line is standing in for.
The next article looks at how foundation models came to exist, and why this shift in AI happened when it did.
References
Bommasani, Rishi, Drew A. Hudson, Ehsan Adeli, et al. (2021), On the Opportunities and Risks of Foundation Models, arXiv:2108.07258, Stanford, CA: Center for Research on Foundation Models, Stanford University, accessed 8 September 2026.
Devlin, Jacob, Ming-Wei Chang, Kenton Lee, et al. (2018), BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, arXiv:1810.04805, accessed 8 September 2026.
European Commission (2025), General-Purpose AI Obligations under the AI Act, last updated 1 August, Brussels: European Commission, accessed 8 September 2026.
Nayak, Pandu (2019), Understanding Searches Better Than Ever Before, The Keyword, 25 October, Mountain View, CA: Google, accessed 8 September 2026.
Ramesh, Aditya, Mikhail Pavlov, Gabriel Goh, et al. (2021), Zero-Shot Text-to-Image Generation, arXiv:2102.12092, accessed 8 September 2026.
Legal instruments
California (2025), Senate Bill 53: Transparency in Frontier Artificial Intelligence Act, Chapter 138, Statutes of 2025, Sacramento, CA: California State Legislature, accessed 8 September 2026.
European Commission (2025), Guidelines on the Scope of the Obligations for Providers of General-Purpose AI Models Established by Regulation (EU) 2024/1689 (AI Act), C(2025) 7719 final, Brussels: European Commission, accessed 8 September 2026.
European Parliament (2023), Amendments Adopted by the European Parliament on 14 June 2023 on the Proposal for a Regulation … Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), P9_TA(2023)0236, Brussels: European Parliament, accessed 8 September 2026.
European Union (2024), Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), Official Journal of the European Union, L series, 12 July, accessed 8 September 2026.
United States (2023), Executive Order 14110 of October 30, 2023: Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, 88 Federal Register 75191, Washington, DC: Office of the Federal Register, accessed 8 September 2026.
United States (2025), Executive Order 14148 of January 20, 2025: Initial Rescissions of Harmful Executive Orders and Actions, 90 Federal Register 8237, Washington, DC: Office of the Federal Register, accessed 8 September 2026.
