Stanford's foundation models report names the category

Stanford’s newly formed Center for Research on Foundation Models, led by Percy Liang, published “On the Opportunities and Risks of Foundation Models,” a roughly 200-page report with well over 100 contributing authors. It proposed and defined the term “foundation model” to describe systems such as GPT-3, BERT and CLIP: models trained on broad data at scale that could be adapted, typically by fine-tuning or prompting, to a wide range of downstream tasks rather than being built and trained separately for each one.

The report’s central argument was that this shared base changed the nature of the underlying technology’s risks and importance. Because a single foundation model could be adapted into many deployed products, defects, biases or failure modes present in the base model would propagate into everything built on top of it, concentrating both capability and risk in ways that earlier, task-specific machine learning systems did not. The report also drew attention to “emergent” capabilities — abilities that appeared unpredictably at larger scale without being explicitly trained for — as a property the field understood poorly, and it surveyed technical, legal and ethical dimensions including data governance, environmental cost, robustness, and downstream inequity.

The report drew criticism as well as attention. Some researchers, including Timnit Gebru, argued that adopting industry-friendly terminology for systems whose harms — including those documented in the “Stochastic Parrots” paper the previous year — were already known amounted to legitimising a technology path rather than critiquing it neutrally. Regardless, “foundation model” became the standard term across research papers, corporate announcements and subsequent regulation, including the EU AI Act, which added specific obligations for general-purpose and foundation models during its later drafting.