Newsfeed / People / Yejin Choi
Yejin Choi

Yejin Choi

researchreasoningstanfordacademic

About Yejin Choi

Yejin Choi is a professor at Stanford University and a MacArthur Fellowship recipient. Her research focuses on natural language processing, commonsense reasoning, and the fundamental limitations of current AI approaches. She is known for her provocative and rigorous critiques of AI hype.

Research Focus

Choi's current work examines:

  • Reasoning in AI - What models can and cannot actually reason about
  • Reinforcement Learning - Why RL isn't the magic bullet many claim
  • Commonsense AI - Teaching machines human-like common sense
  • AI Limitations - Understanding what current approaches fundamentally cannot do

Notable Positions

On Reinforcement Learning

Choi argues that despite the excitement around RL for reasoning, it ultimately reduces to data:

"This sounds more magical but under the hood this all boils down to synthesizing even more data because that much data is still not good enough and we somehow need to now ask AI to generate more data."

She notes that with enough effort, supervised fine-tuning (like OpenThought) can beat RL approaches, suggesting RL isn't fundamentally superior.

On Abductive Reasoning

Choi highlights a crucial gap in AI reasoning capabilities. While models do inductive and deductive reasoning, they struggle with abductive reasoning - forming hypotheses from partial observations:

"Most reasoning - induction and deduction - is regurgitation of the same information that you already had. Whereas abduction is this mental act of coming up with the best possible explanation of your partial observation."

She famously points out that Sherlock Holmes' "deductions" are actually abductions - creative leaps to the best explanation.

On AI's Fundamental Limitations

Choi's most profound critique is that current AI is limited to "the neighborhood of internet data":

"Human knowledge is not equivalent to the universe of actual knowledge out there. There are truths about how to cure cancer that are not on the internet. The current way of doing curation of data isn't going to really teach the model how to come up with personalized drugs that could prevent or cure cancer."

Key Quotes

  • "It all comes back to data."
  • "Abduction is this mental act of coming up with the best possible explanation of your partial observation."
  • "Induction and deduction is regurgitation of the same information that you already had."