Technical

Foundation Models

Large-scale models trained on broad data for many tasks.

How the laws define this

Regulators and official guidance agree that foundation models are large-scale AI models trained on broad, diverse, or general-purpose data that can be adapted or fine-tuned for a wide range of downstream tasks. The main point of divergence among these instruments lies in their regulatory focus and practical implications. The Artificial Intelligence Act treats these models with specific model-level obligations—including transparency, monitoring, and mitigation measures—when system-level risk arises. In contrast, the Digitaliseringsstyrelsen Guides for Responsible Use of Generative AI focus on adaptation risks alongside procurement and governance implications. Meanwhile, the Model AI Governance Framework for Generative AI specifically highlights generative applications as downstream uses, whereas the G7 Hiroshima AI Process: International Guiding Principles and Code of Conduct emphasizes general-purpose capabilities and broad applicability without detailing specific regulatory obligations or procurement measures.

Synthesised from the 4 statutory definitions below. Each one is quoted in full, with its source.

How jurisdictions differ

What is distinctive in each definition, binding law in force first. Showing 3 of the 4 instruments that define this term; all are listed under Definitions below.

Jurisdiction	Instrument	Status	What is distinctive
Denmark	Digitaliseringsstyrelsen Guides for Responsible Use of Generative AI	In Force	Highlights generality, adaptation risks, and implications for procurement and governance in guides for Danish public authorities and private enterprises.
G7	Hiroshima Process International Guiding Principles for Organizations Developing Advanced AI Systems and International Code of Conduct for Organizations Developing Advanced AI Systems	Adopted	Characterizes them by general-purpose capabilities and broad applicability in the G7 international framework for advanced AI developers.
Norway	Artificial Intelligence Act	Draft	Specifies model-level obligations like transparency, monitoring, and mitigation when system-level risk arises under the Norwegian draft law.

What it triggers

Under the Norwegian Artificial Intelligence Act draft, these models would face specific model-level obligations including transparency, monitoring, and mitigation measures when system-level risk arises.source

Definitions (4)

Foundation Models are large-scale AI models trained on broad and diverse datasets that can be adapted or fine-tuned for a wide range of downstream tasks, characterized by general-purpose capabilities and broad applicability.

G7 Hiroshima AI Process: International Guiding Principles and Code of ConductDefinition 1 of 4

Large-scale models trained on broad, diverse datasets that can be fine-tuned or adapted for many downstream tasks and use-cases; the guides highlight their generality, adaptation risks, and implications for procurement and governance.

Digitaliseringsstyrelsen Guides for Responsible Use of Generative AI (Guides til ansvarlig anvendelse af generativ kunstig intelligens)Definition 2 of 4

Models trained on broad, general-purpose data that can be adapted for many downstream tasks (foundation or general‑purpose models); the draft law treats these with specific model‑level obligations when system‑level risk arises, including transparency, monitoring, and mitigation measures.

Artificial Intelligence ActDefinition 3 of 4

Large-scale AI models trained on vast and diverse datasets that serve as a base for a wide range of downstream tasks, including generative applications, and can be fine-tuned or adapted for specific uses.

Model AI Governance Framework for Generative AIDefinition 4 of 4

Not to be confused with

Foundation models / Frontier AI

Refers to highly capable general-purpose models often termed frontier AI, while this term focuses on training data and adaptability without emphasizing extreme capability.

Large/fundamental models

Limits to models with at least one billion parameters, whereas this term lacks any parameter size threshold.

Questions and answers

What obligations apply to foundation models in Norway?

The Norwegian Artificial Intelligence Act draft treats foundation models with specific model-level obligations when system-level risk arises. These include transparency, monitoring, and mitigation measures. It adopts a risk-based approach similar to the EU AI Act.

How does the G7 define foundation models?

The G7 Hiroshima Process defines them as large-scale AI models trained on broad and diverse datasets that can be adapted for a wide range of downstream tasks. They are characterized by general-purpose capabilities and broad applicability.

What risks do Danish guides highlight for foundation models?

The Digitaliseringsstyrelsen Guides highlight generality, adaptation risks, and implications for procurement and governance of foundation models. They guide Danish public authorities and private enterprises on risk management.