A large AI model pre-trained on massive data, designed to be adapted to many downstream tasks, the source of a common dependence at the scale of the economy.
A foundation model is a very large artificial intelligence model, trained on massive volumes of data and designed not for a single task but to serve as an adaptable base for a multitude of downstream applications, from dialogue to code writing to image analysis. The term was popularized in 2021 by researchers at Stanford, who underlined both the power and the risks of this paradigm shift. The power lies in risk pooling, since a single pre-trained model can be reused and specialized at low cost for countless uses. The risk lies in the concentration this pooling creates, because when thousands of applications rest on a small number of foundation models, they inherit their biases, their flaws and their blind spots, and a failure or vulnerability of the base model propagates to the whole ecosystem. This is precisely the mechanism of algorithmic monoculture, a single point of failure that is not technical but cognitive. For insurance, this common dependence creates an accumulation exposure of a new kind, in which apparently independent risks in fact share the same underlying brain.
Dozens of different legal-assistance tools rely on the same foundation model. A reasoning flaw common to that model can then produce the same error across all those tools at once, creating an accumulation no one had identified.
foundation model, modèle socle, modèle de base