The term was coined by Stanford researchers in 2021 to describe a new paradigm in AI development: rather than training specialized models for specific tasks, organizations train or license a single large model and adapt it for many uses. GPT-4, Claude, Gemini, and Llama are all foundation models.
Foundation models derive their power from scale: they are trained on datasets large enough to capture broad knowledge and language capabilities, and their general-purpose nature makes them highly adaptable. A single foundation model can, with appropriate prompting or fine-tuning, serve as a coding assistant, a customer service agent, a document summarizer, or a data analyst.
For enterprise data governance, foundation models present specific challenges. When a foundation model is fine-tuned on enterprise data, that data becomes embedded in the model's weights, potentially in ways that are difficult to audit, attribute, or reverse. This makes the 'can I use this data?' question critical to answer before fine-tuning begins, not after.