Yejin Choi

wiki
, , , ,
1

Korean-American NLP researcher and MacArthur Fellow whose career has traced the arc from opinion mining and image captioning to commonsense knowledge bases, moral reasoning AI, and the case for bending scaling laws with smarter algorithms — now at Stanford as the Dieter Schwarz Foundation Professor.

Profile
Born 1977, South Korea
Nationality Korean-American
Current Institution(s) Stanford University — Dieter Schwarz Foundation Professor & Senior Fellow, Computer Science and HAI (2025–present)
Research Areas Natural Language Processing, Commonsense AI, Moral Reasoning, Neuro-Symbolic AI, Pluralistic Alignment, AI for Science, Language Model Foundations
Doctoral Advisor Claire Cardie
Doctoral Thesis Fine-Grained Opinion Analysis: Structure-Aware Approaches (Cornell University, 2010)
Website yejinc.github.io
X / Twitter @YejinChoinka
GitHub yejinc
Google Scholar Yejin Choi

Overview

Yejin Choi (Korean: 최예진) is a Korean-American NLP researcher and the 2022 MacArthur Fellow who is now the Dieter Schwarz Foundation Professor at Stanford University and a Senior Fellow at the Stanford Institute for Human-Centered Artificial Intelligence (HAI). Her career is unusually broad: she moved from opinion mining and sentiment analysis (her doctoral work at Cornell) to image captioning and visual language (earning the CVPR Longuet-Higgins Prize and ICCV Marr Prize), to commonsense reasoning and knowledge bases (ATOMIC, COMET, Winogrande, at the University of Washington and the Allen Institute for AI), to moral AI systems (Delphi), to research on the limits and failures of large language models, to AI for science at NVIDIA, and most recently to pluralistic alignment and the thesis that “brighter algorithms” can transcend scaling laws. She has received best or outstanding paper awards at more than fifteen major venues across NLP, vision, and ML, given a TED talk watched by millions, briefed the United Nations Security Council on AI safety in September 2025, and is named among TIME100 AI in both 2023 and 2025. The breadth of her work is matched by her public profile: no active NLP researcher has communicated the gap between AI’s apparent capabilities and its underlying limitations — what she calls its being “incredibly smart and shockingly stupid” — to a wider mainstream audience.

Early Life & Education

Choi was born in South Korea in 1977. She completed a bachelor’s degree in Computer Science at Seoul National University before moving to the United States for graduate study. At Cornell University, she worked with Claire Cardie in natural language processing, completing her PhD in 2010 with a dissertation titled Fine-Grained Opinion Analysis: Structure-Aware Approaches. The thesis developed structured probabilistic models for identifying opinion holders, targets, and sentiment polarity in text — an early contribution to what became the field of aspect-based sentiment analysis. Her work on deceptive opinion identification (including fake hotel review detection, published at ACL 2011 with Myle Ott, Claire Cardie, and Jeffrey Hancock) earned the ACL 2021 Test of Time Award, recognizing it as a foundational result in computational deception detection.

Career

Stony Brook University — Assistant Professor (2010–2013)

After completing her PhD, Choi joined the Computer Science Department at Stony Brook University as an assistant professor. During this period she extended her NLP work to include visual language grounding. Her BabyTalk project — generating natural language descriptions of images — was an early demonstration of automatic image captioning, published in IEEE TPAMI. Work from this period was recognized with the CVPR 2021 Longuet-Higgins Prize (for its lasting influence on the field) and the ICCV 2013 Marr Prize (best paper), the latter for work on learning entry-level categories from large-scale image datasets with Alexander Berg’s group.

University of Washington — Professor (2013–2024)

Choi joined the Paul G. Allen School of Computer Science & Engineering at the University of Washington, where she would spend over a decade as one of its most prominent NLP researchers. She held the Brett Helsel Career Development Professorship (2020–2023) and the Wissner-Slivka Chair (2023–2024). During this period she co-led the Mosaic group at the Allen Institute for AI (AI2), embedded at UW with AI2 researchers.

Commonsense AI: ATOMIC and COMET. The most sustained intellectual commitment of Choi’s UW years was the question of whether AI systems could acquire and reason with commonsense knowledge — the implicit understanding of everyday events, causality, and social norms that humans take for granted but that NLP systems consistently lacked. She assembled ATOMIC (Atlas of Machine Commonsense), a knowledge graph of over 300,000 if-then inferential relationships covering causes, effects, intents, and reactions of everyday events (published at AAAI 2019, Outstanding Paper Award 2020). ATOMIC became the largest open commonsense knowledge base of its kind. Building on it, she developed COMET (Commonsense Transformers), showing that large language models fine-tuned on ATOMIC could generate new commonsense inferences in natural language — an early demonstration of LLMs as implicit knowledge base completion systems. The program was covered by Quanta Magazine (“Common Sense Comes to Computers”) and Axios.

Winogrande. Choi led the construction of Winogrande, a large-scale crowdsourced dataset for the Winograd schema challenge — a class of pronoun disambiguation problems specifically designed to require commonsense reasoning. Winogrande (NeurIPS 2020) used an adversarial filtering algorithm to remove data artifacts that allow models to cheat, creating a substantially harder benchmark than previous versions. The dataset became a standard evaluation for commonsense reasoning in language models.

Delphi and moral AI. Choi’s group developed Delphi, a computational model of moral judgments (2021), and the associated Commonsense Norm Bank dataset. Delphi was one of the first large-scale attempts to train a model on explicit moral intuitions across diverse social situations, generating human-readable moral assessments. The project was featured by The New York Times (“Can a Machine Learn Morality?”) and sparked sustained public discussion about the role of AI in ethical decision-making. Choi has maintained a nuanced position: Delphi was designed as a research tool for understanding normative reasoning in language models, not as a prescription for how AI should make moral decisions.

Alexa Prize. The UW team led by Choi (the Sounding Board project) won the inaugural Amazon Alexa Prize for developing the most engaging open-domain conversational AI among university teams.

AI limitations and “shockingly stupid” AI. In parallel with her constructive commonsense work, Choi developed a public intellectual voice around the gap between the surface capabilities of large language models and their underlying understanding. Her 2023 TED Talk, “Why AI is Incredibly Smart — and Shockingly Stupid,” became widely viewed and is among the most-shared academic contributions to public understanding of LLM limitations.

Allen Institute for AI (AI2) — Senior Research Manager, Mosaic Group (2018–2024)

In 2018 Choi joined AI2 as a senior researcher embedded at UW, co-leading the Mosaic group — the commonsense AI research program that produced ATOMIC, COMET, Winogrande, and related knowledge and reasoning resources. She maintained a joint appointment there throughout her UW tenure.

NVIDIA — Distinguished Scientist (2024–2025)

Before joining Stanford, Choi held a Distinguished Scientist role at NVIDIA, where her work expanded into AI for science — molecular foundation models, protein reasoning, crystal structure prediction — and contributed to the Nemotron model family. She also developed work on bending scaling laws with “brighter algorithms”: the thesis that advances in algorithm design, training strategies, and data curation (rather than raw compute scaling) can unlock qualitative capability improvements in small and medium models. Papers from this period include work on prolonged reinforcement learning, test-time training, and new tokenization approaches.

Stanford University (2025–present)

In 2025, Stanford HAI announced Choi’s appointment as the Dieter Schwarz Foundation Professor of Computer Science and a Senior Fellow at HAI. Her research interests at Stanford span AI for science (molecular and protein reasoning), pluralistic alignment and values in AI, alternative training and inference algorithms, and the study of LLM limitations. In September 2025 she briefed the United Nations Security Council on AI safety, one of the most prominent public-policy invitations any AI researcher has received.

Key Contributions

  • Fake Review Detection (ACL 2011; ACL 2021 Test of Time Award) — “Finding Deceptive Opinion Spam by Any Stretch of the Imagination,” with Myle Ott, Claire Cardie, and Jeffrey Hancock. Introduced the first large-scale study of deceptive opinion generation and detection, demonstrating that machine learning could identify fake hotel reviews with above-chance accuracy. Received the ACL Test of Time Award a decade later for foundational impact on computational credibility research.

  • BabyTalk / Image Captioning (CVPR Longuet-Higgins Prize 2021) — “BabyTalk: Understanding and Generating Simple Image Descriptions,” TPAMI 2013, with Girish Kulkarni, Vicente Ordonez, and the Berg group. One of the earliest automated natural language description systems for images, earning retroactive recognition as a foundational contribution to the vision-language grounding field.

  • Entry-Level Categories (ICCV 2013 Marr Prize) — Work on learning the level of abstraction at which humans naturally name objects in images, winning the top paper award at ICCV 2013.

  • ATOMIC (AAAI 2019; AAAI 2020 Outstanding Paper) — Atlas of Machine Commonsense: a knowledge graph of over 300,000 if-then inferential rules covering causes, effects, intents, and reactions of everyday events. The largest open commonsense knowledge base of its time, forming the basis for COMET and multiple generations of commonsense reasoning systems.

  • COMET (ACL 2019) — Commonsense Transformers: fine-tuned language models that generate new commonsense inferences from ATOMIC, demonstrating LLMs as generative knowledge base completion systems and as commonsense inference engines.

  • Winogrande (NeurIPS 2020) — A large-scale adversarially filtered crowdsourced dataset for Winograd schema commonsense reasoning, substantially harder than earlier versions and widely adopted as a standard benchmark for measuring commonsense understanding in language models.

  • Delphi (2021) — A computational model of moral judgment trained on the Commonsense Norm Bank dataset, one of the first attempts to create a language model that produces human-readable moral assessments across diverse social situations. Featured in The New York Times and contributed to broader public debate about AI and ethics.

  • Pluralistic Alignment and Bending Scaling Laws — A sustained research agenda (2023–present) arguing that diverse human values and perspectives must be explicitly represented in AI training rather than averaged away, combined with a technical program demonstrating that intelligent algorithm design can achieve capability improvements without simply scaling compute.

Awards & Recognition

  • MacArthur Fellowship (2022) — The “genius grant” from the MacArthur Foundation, awarded for exceptional creativity and potential.
  • ACL Fellow (2022) — Fellow of the Association for Computational Linguistics.
  • TIME100 AI (2025 and 2023) — Named to Time magazine’s list of the 100 most influential people in AI in both years.
  • AI2050 Senior Fellow (2024) — Named by Schmidt Sciences’ AI2050 program.
  • UN Security Council Briefing (September 2025) — Briefed the UN Security Council on AI safety.
  • Best/Outstanding Paper Awards — At NeurIPS 2025 (two awards), EMNLP 2025 (two awards), ACL 2025 (two awards), EMNLP 2023, ACL 2023 (two awards), NAACL 2022, ICML 2022, NeurIPS 2021, AAAI 2020.
  • CVPR Longuet-Higgins Prize (2021) — For the BabyTalk image captioning work.
  • ACL Test of Time Award (2021) — For the fake review detection paper.
  • ICCV 2013 Marr Prize — Best paper award at the International Conference on Computer Vision.
  • Anita Borg Early Career Award (BECA, 2018) — CRA-WP award for outstanding early-career women in computing.
  • IEEE AI’s 10 to Watch (2016) — Named among the ten most influential early-career AI researchers by IEEE.
  • Alexa Prize Winner — University of Washington’s Sounding Board team (led by Choi) won the inaugural Amazon Alexa Prize.

Key Relationships

  • Claire Cardie — PhD advisor at Cornell; a leading researcher in information extraction and opinion mining who shaped Choi’s foundational NLP orientation.
  • Hannaneh Hajishirzi — Long-term University of Washington colleague; frequent co-author on language model foundations, open models (OLMo, OLMoTrace), and reading comprehension.
  • Luke Zettlemoyer — University of Washington colleague; co-authored on multiple NLP benchmarks, language model analysis, and grounding research.
  • Maarten Sap — PhD student who co-authored ATOMIC, COMET, and multiple moral/commonsense AI papers; now assistant professor at CMU.
  • Ronan Le Bras — Long-running collaborator at AI2 on commonsense resources and reasoning benchmarks.
  • Liwei Jiang — PhD student at UW/Stanford working on pluralistic alignment, values in AI, and normative reasoning.
  • Ximing Lu — Close collaborator and researcher on reasoning, RLVR, and model training; appeared across dozens of recent papers.
  • Noah A. Smith — UW professor and frequent collaborator on language model evaluation, data, and NLP foundations.

Personal Style

Choi is one of the most public-facing academic AI researchers of her generation — not primarily through social media, but through a consistent effort to communicate the nuanced reality of AI capabilities to non-specialist audiences. Her TED Talk, New Yorker profile, Quanta Magazine coverage, This American Life appearance, and Bill Gates podcast have collectively reached audiences in the tens of millions, and she approaches these opportunities with the same intellectual precision she brings to technical work: resisting both AI hype and AI dismissiveness in favor of precise articulation of what current systems can and cannot do. Her research agenda is characteristically ambitious in scope — commonsense AI, moral reasoning, pluralistic alignment, and AI for science are each sufficiently large to anchor an entire research program, and she pursues them all simultaneously. She has spoken about the importance of representation in AI, about bias in language and vision systems that disadvantages underrepresented groups, and about the obligation of AI researchers to participate in public policy discourse. Her September 2025 UN Security Council briefing represents the most direct intersection of her research with international governance of the technology she studies.

References