Foundation model

The term was coined by Stanford researchers (Bommasani et al.) to describe models such as large language models and multimodal image and text models that serve as a common base for many applications. A single foundation model can be fine-tuned, prompted or connected to other components to write text, summarize documents, generate images, code or classify content. That reuse is the source of both their value and their risk: a flaw in the base model, such as bias, a tendency to hallucinate or a security weakness, propagates to every system built on it.

The EU AI Act does not use the term; it regulates general-purpose AI models, a closely overlapping legal category, with documentation, copyright policy and training content summary duties, plus additional obligations for models with systemic risk. Organizations building on a foundation model remain responsible for evaluating it in their own context and for the obligations of their role in the value chain.

Example: a startup builds a contract review product on a commercial large language model and runs its own accuracy tests on legal documents in its market.