Preprint
Machine Learning

On the opportunities and risks of foundation models

Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Ré, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, Percy Liang
January 1, 2021
arXivPDF

0

Citations

0

Influential Citations

—

Venue

2021

Year

Abstract

… Though foundation models are based on standard deep … widespread deployment of foundation models, we currently lack … of the critical research on foundation models will require deep …

Analysis

Why This Paper Matters

This paper is a seminal work that introduced the term "foundation models" to the AI community, capturing the paradigm shift towards large-scale pre-trained models like BERT, GPT-3, and CLIP. It matters because it systematically articulates both the immense potential and the profound risks of these models, which have become central to modern AI. The paper serves as a wake-up call for researchers and practitioners to consider not just performance gains but also ethical, environmental, and societal consequences.

The significance lies in its balanced perspective: while foundation models enable unprecedented capabilities in natural language processing, computer vision, and multimodal tasks, they also inherit and amplify biases from training data, pose challenges for interpretability, and require massive computational resources. This dual framing has shaped subsequent research agendas and policy discussions.

Technical Contributions

  • Definition and taxonomy: The paper clearly defines foundation models as models trained on broad data at scale that can be adapted to a wide range of downstream tasks.
  • Identification of emergent properties: It highlights how scaling leads to emergent abilities not present in smaller models, such as in-context learning and reasoning.
  • Risk categorization: The authors systematically categorize risks into technical (e.g., bias, robustness), societal (e.g., misuse, economic impact), and environmental (e.g., carbon footprint).
  • Call for interdisciplinary research: The paper argues that addressing these challenges requires collaboration across machine learning, ethics, law, and social sciences.

Results

As a position paper, it does not present experimental results. Instead, it synthesizes observations from existing models (e.g., GPT-3, BERT) to illustrate points. For example, it notes that GPT-3 exhibits biases related to race and gender, and that training a single large model can emit as much carbon as several cars over their lifetimes. These qualitative findings have been corroborated by later empirical studies.

Significance

The paper has had a lasting impact on the AI field by coining a term that unified research on large pre-trained models. It has been cited extensively and has influenced both academic research and industry practices. It prompted initiatives like the Stanford Center for Research on Foundation Models (CRFM) and shaped discussions around responsible AI development. The paper's balanced view encourages practitioners to innovate while remaining vigilant about the broader implications of their work.

foundation-modelsdeep-learningnlpai-ethicsbiasscalingtransfer-learning

Related Papers