Foundation Model
Basismodell; in the EU AI Act: KI-Modell mit allgemeinem Verwendungszweck (general-purpose AI model)
What it means
A foundation model is a large AI model trained on broad data so that it can be adapted to many different tasks, rather than being built for one purpose. Large language models are the best-known type, but foundation models also exist for images, speech and other data.
Real-world example
A national archive and a city health department both build assistants on the same commercially available foundation model. The archive connects it to its finding aids to help researchers locate records; the health department connects it to its internal guidance so staff can find procedures. The underlying model is the same, but the data, safeguards and risks are very different.
Why it matters for AI adoption
Most organisations will not train their own foundation model; they will build on one from a provider. That makes the choice of model and provider a strategic dependency affecting cost, performance, data protection and long-term flexibility. The EU AI Act sets specific obligations for providers of general-purpose AI models.
Questions to ask
- Which foundation model is this built on, and could we switch to another if needed?
- What does the provider disclose about the model's training, limitations and evaluations?