What’s in the Chatterbox? Large Language Models, Why They Matter, and What We Should Do About Them

April 2022
Johanna Okerlund, Evan Klasky, Aditya Middha, Sujin Kim, Hannah Rosenfeld, Molly Kleinman, Shobita Parthasarathy

Large language models (LLMs)—machine learning algorithms that can recognize, summarize, translate, predict, and generate human languages on the basis of very large text-based datasets—are likely to provide the most convincing computer-generated imitation of human language yet. Because language generated by LLMs will be more sophisticated and human-like than their predecessors, and because they perform better on tasks for which they have not been explicitly trained, we expect that they will be widely used. Policymakers might use them to assess public sentiment about pending legislation, patients could summarize and evaluate the state of biomedical knowledge to empower their interactions with healthcare professionals, and scientists could translate research findings across languages. In sum, LLMs have the potential to transform how and with whom we communicate.

This report first summarizes the LLM landscape and the technology’s basic features. We then outline the implications identified through our ACS approach. We conclude that LLMs will produce enormous social change including:

  1. Exacerbating environmental injustice
  2. Accelerating the thirst for data
  3. Becoming quickly integrated into existing infrastructure
  4. Reinforcing inequality
  5. Reorganizing labor and expertise
  6. Increasing social fragmentation.

LLMs will transform a range of sectors, but the final section of the report focuses on how these changes could unfold in one specific area: scientific research. Finally, using these insights we provide informed guidance on how to develop, manage, and govern LLMs.