- https://arxiv.org/pdf/2108.07258
- Source: arxiv
TL;DR
The Stanford CRFM report that coined the term “foundation model” systematically analyzes opportunities across healthcare, law, and science alongside risks including bias amplification, economic concentration, and environmental cost, arguing the paradigm demands coordinated research on alignment, robustness, and governance.
Summary
Large models trained on broad data and adapted to many downstream tasks represent a new AI paradigm, which this report names foundation models. It maps opportunities across healthcare, law, education, and scientific discovery, and identifies risks: bias amplification, misuse potential, and economic power concentration when a small number of providers control the underlying models. The research agenda calls for coordinated progress on alignment, robustness, interpretability, and governance — framing these as prerequisites, not afterthoughts, for societal benefit.
Key contributions
- Coins and defines the term “foundation model”, establishing a vocabulary for the large-scale pre-training paradigm.
- Provides a systematic mapping of opportunity domains (healthcare, law, education, science) against concrete risks.
- Identifies bias amplification as a structural risk when a single model underlies many downstream applications.
- Frames governance and policy coordination as necessary complements to technical alignment research.
- Establishes interpretability and environmental cost as first-class research priorities alongside capability.
When to cite
- When defining or introducing the concept of foundation models and justifying the terminology.
- When discussing bias and fairness risks arising from widespread deployment of a single pre-trained model.
- When arguing for interpretability research as a prerequisite to safe foundation model deployment.
- When framing AI governance requirements around large-scale pre-trained models.