Yejin Choi
Last reviewed
May 31, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 3,724 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 3,724 words
Add missing citations, update stale details, or suggest a clearer explanation.
Yejin Choi (born 1977) is a South Korean–American computer scientist and the Dieter Schwarz Foundation HAI Professor and Professor of Computer Science at stanford university, where she is also a Senior Fellow at the Stanford Institute for Human-Centered Artificial Intelligence (Stanford HAI).[1][2] She is best known for shaping the modern research agenda on commonsense reasoning, social and moral reasoning, and constrained generation in natural language processing, and for advocating that smaller, knowledge-rich language models can rival much larger systems trained at greater compute.[1][3][4]
Before joining Stanford in 2025, Choi spent a decade on the faculty of the Paul G. Allen School of Computer Science & Engineering at the University of Washington, where she held the Brett Helsel Career Development Professorship and, in 2023, the Wissner-Slivka Chair, with a joint appointment as Senior Research Manager at the allen institute for ai (also known as ai2).[3][5][6] In August 2024 she left both UW and AI2 to move to California, joining nvidia as Senior Director of AI Research before transitioning to Stanford the following year.[7][2]
Choi was named a MacArthur Fellow in 2022 for using natural language processing to develop AI systems capable of commonsense reasoning, and she has been named to TIME's list of the 100 most influential people in AI in both 2023 and 2025.[8][9][4] Her 2023 TED talk, "Why AI is incredibly smart and shockingly stupid," has become one of the most widely cited public explanations of the limits of large language models.[10][11]
| Born | 1977, South Korea[12] |
| Education | BS, Seoul National University (1999); PhD in Computer Science, Cornell University (2010, advised by Claire Cardie)[8][13][12] |
| Current position | Dieter Schwarz Foundation HAI Professor and Professor of Computer Science, stanford university; Senior Fellow, Stanford HAI (2025–present)[1][2] |
| Previous positions | Assistant Professor, SUNY Stony Brook (2010–2014); Assistant → Associate → Full Professor, University of Washington (2014–2024); Senior Research Manager, allen institute for ai (2018–2024); Senior Director of AI Research, nvidia (2024–2025)[12][5][7] |
| Best-known projects | ATOMIC, COMET, Delphi, Social IQa, Social Chemistry 101, hellaswag, winogrande, toxigen (collaborator), NeuroLogic Decoding, Impossible Distillation[14][15][16][17][18][19][20][21][22] |
| Notable honors | MacArthur Fellowship (2022); ACL Fellow (2022); AI2050 Senior Fellow (2024); TIME100 AI (2023, 2025); ACL Test-of-Time Award (2021); cvpr Longuet-Higgins Prize (2021); iccv Marr Prize (2013); Borg Early Career Award (2018); IEEE AI's 10 to Watch (2016)[8][4][23][24] |
| Public profile | TED 2023 main-stage speaker; subject of features in The New York Times, NPR, GeekWire, and Time[10][25][26] |
Yejin Choi was born in 1977 in South Korea and grew up there before moving to the United States for graduate study.[12] She earned a Bachelor of Science in computer science and engineering from Seoul National University in 1999, then enrolled at Cornell University in upstate New York for her doctorate.[8][12]
At Cornell, Choi worked with the natural language processing researcher Claire Cardie. Her dissertation, "Fine-Grained Opinion Analysis: Structure-Aware Approaches," focused on extracting opinions and their sources from text using structured prediction methods.[13][12] Among the papers from her doctoral period, "Identifying sources of opinions with conditional random fields and extraction patterns," co-authored with Cardie and others in 2005, became an early standard reference in opinion extraction.[12] She received her PhD in computer science from Cornell in 2010.[8][13]
After completing her PhD, Choi joined the State University of New York at Stony Brook as an assistant professor of computer science. Her early work at Stony Brook, between 2010 and 2014, applied NLP to detect deception in online text. The paper "Finding Deceptive Opinion Spam by Any Stretch of the Imagination," presented in 2011, used computational linguistics to distinguish fake from authentic hotel reviews and helped launch a strand of research on text-based deception detection.[12]
In 2014, Choi moved across the country to the Paul G. Allen School of Computer Science & Engineering at the University of Washington in Seattle, where she rose through the ranks to full professor.[5] At UW she helped lead a UW Natural Language Processing group that during the late 2010s and early 2020s produced some of the most widely used commonsense reasoning benchmarks and language-generation methods in the field.
In 2017, Choi was a faculty advisor for the UW team Sounding Board, a socialbot that won the inaugural Amazon Alexa Prize, beating teams from more than twenty countries to a US$500,000 first-place award.[27]
She joined the allen institute for ai, the Seattle-based research nonprofit founded by Microsoft co-founder Paul Allen, in 2018, eventually serving as senior research manager and leading the AI2 Mosaic team, the institute's group focused on commonsense reasoning and language understanding.[12][28]
At UW, Choi held the Brett Helsel Career Development Professorship from 2020 to 2023 and was named to the endowed Wissner-Slivka Chair in 2023.[6][29] Over the course of her decade in Seattle she received an exceptionally large concentration of "best paper" and "outstanding paper" recognitions at top venues, including AAAI 2020, NeurIPS 2021, the ACL Test-of-Time Award in 2021, the cvpr Longuet-Higgins Prize in 2021, NAACL Best Paper in 2022, ICML Outstanding Paper in 2022, ACL Best Paper in 2023, and EMNLP Outstanding Paper in 2023.[4][29]
In August 2024 Choi announced internally at AI2 and the Allen School that she was leaving both organizations. "Even though everything is so wonderful here, I am adventure-seeking by nature," she wrote to UW colleagues, indicating that she planned to move to California, with a new role to be confirmed.[7] Her next position was as Senior Director of AI Research at nvidia, based in the company's research organization.[7][2]
On 16 January 2025, Stanford HAI announced that Choi would join the institute as the inaugural Dieter Schwarz Foundation HAI Professor, with a full professorship in the Stanford Department of Computer Science and a Senior Fellowship at Stanford HAI.[2] Stanford HAI's announcement positioned the appointment as central to its research agenda on aligning AI with societal values, with a particular emphasis on Choi's ongoing work on common-sense reasoning and on the shift from large language models to smaller, more efficient language models.[2][1]
The Dieter Schwarz Foundation chair at Stanford is an endowed professorship funded by the German Dieter Schwarz Foundation; Choi is the inaugural holder of the position at Stanford HAI.[2] Her Stanford profile describes her research themes as common-sense intelligence, alternative training methods including test-time training and reinforced pretraining, pluralistic alignment, theory of mind, and AI for science such as molecular foundation models and protein reasoning.[1]
A central focus of Choi's career has been giving AI systems "common sense," the kind of implicit world knowledge that humans deploy effortlessly but which had long eluded NLP systems.[8] She has argued that purely scaling up language models on web data is insufficient, because commonsense knowledge is often unstated in text and is interwoven with norms and values that web data does not represent reliably.[11]
In 2019 Choi was senior author on ATOMIC ("An Atlas of Machine Commonsense for If-Then Reasoning"), a knowledge graph of approximately 877,000 textual descriptions of inferential, causal, and social commonsense relations (e.g., if X pays Y a compliment, Y will likely return the compliment), structured around nine if-then relation types.[14] ATOMIC was followed by COMET (Commonsense Transformers), introduced at ACL 2019, in which a neural language model was trained on commonsense knowledge graphs so that it could generate commonsense inferences for previously unseen situations rather than retrieve them, with human judges rating more than 75% of its outputs as plausible.[8][14] A 2021 successor, COMET-ATOMIC 2020, scaled and combined symbolic and neural commonsense knowledge graphs.[30]
Choi was also a co-author of several other widely used commonsense and reasoning benchmarks created during her UW/AI2 years, including hellaswag (Zellers, Holtzman, Bisk, Farhadi, and Choi, ACL 2019), a benchmark of "adversarial" sentence-completion problems that became a standard evaluation for large language models; winogrande (Sakaguchi, Le Bras, Bhagavatula, and Choi, AAAI 2020), a large-scale Winograd Schema challenge that won an AAAI 2020 Outstanding Paper Award; and Social IQa (Sap, Rashkin, Chen, Le Bras, and Choi, EMNLP-IJCNLP 2019), the first large-scale benchmark for commonsense reasoning about everyday social situations.[15][16][31][17]
The CVPR 2021 Longuet-Higgins Prize, awarded retroactively to her 2013 work on generating image descriptions ("BabyTalk"), recognized her earlier influential contributions at the boundary of language and vision.[12][4]
Starting around 2020, Choi's group extended commonsense modeling into the moral and social domain. Social Chemistry 101 (Forbes, Hwang, Shwartz, Sap, and Choi, EMNLP 2020) introduced a 292,000-entry corpus of "rules of thumb," everyday social and moral norms such as "It is rude to run a blender at 5 a.m.," annotated along twelve dimensions of human judgment.[19] The Neural Norm Transformer model trained on Social Chemistry 101 could generalize to previously unseen situations.[19]
In October 2021, Choi's team at AI2 released Delphi, an experimental neural system designed to make moral judgments about everyday scenarios.[25][32] The accompanying paper, "Can Machines Learn Morality? The Delphi Experiment," reported that human judges rated Delphi's ethical judgments as up to 92% accurate.[32] Delphi became one of the most heavily discussed AI experiments of its time: it drew substantial press coverage in The New York Times, NPR, and elsewhere, but also significant academic criticism from scholars who questioned both the framing of "machine morality" and the methodology of treating ethics as a labeled-data classification problem.[25][32] After early complaints, the project relabeled Delphi as "a research prototype designed to model people's moral judgments" rather than an oracle that "says" what is right or wrong.[25] An updated, peer-reviewed account of the experiment subsequently appeared in Nature Machine Intelligence.[33]
In a related strand, toxigen (Hartvigsen, Gabriel, Palangi, Sap, Ray, and Kamar, ACL 2022), on which Choi's longtime collaborators Sap and Gabriel were authors, introduced a large machine-generated dataset for adversarial and implicit hate-speech detection covering thirteen minority identity groups, and helped catalyse a line of work in Choi's broader network on detoxifying language models.[34]
Choi's group has also worked extensively on the decoding side of generation. NeuroLogic Decoding (Lu, West, Zellers, Le Bras, Bhagavatula, and Choi, NAACL 2021) introduced a beam-search algorithm that lets a neural language model satisfy arbitrary predicate-logic constraints on its output without retraining the underlying model.[20] Follow-up work, including NeuroLogic A*esque decoding, extended this with lookahead heuristics for constrained generation.[20]
This decoding line is closely tied to Choi's interest in what she has called pluralistic alignment: methods and data designed to reflect a plurality of human values rather than collapse to a single voice. Her Stanford profile lists pluralistic alignment, theory of mind, and "alternative training methods" such as test-time training and reinforced pretraining among her current research themes.[1]
A recurring theme of Choi's recent talks and papers is the argument that knowledge- and algorithm-rich smaller models can compete with much larger ones. Impossible Distillation (Jung, West, Jiang, Brahman, Lu, Fisher, Sorensen, and Choi, NAACL 2024) showed that a 770-million-parameter student model trained on a dataset distilled from a smaller and "low-quality" teacher language model could outperform the 175-billion-parameter GPT-3 on summarization and paraphrasing tasks, as judged by both automatic metrics and human evaluators.[21][22] The work also released DIMSUM+, a dataset of 3.4 million machine-generated summaries and paraphrases.[22] Choi's Stanford HAI announcement explicitly described her continuing work on "the shift from LLMs to SLMs" (large to small language models) as a central reason for her recruitment.[2]
In November 2025, a paper led by Choi titled "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)" received a neurips 2025 Best Paper Award. Using a benchmark of 26,000 real-world open-ended queries, the team showed that different leading large language models tend to converge on strikingly similar responses, an "artificial hivemind" effect with implications for diversity, safety, and information ecology.[35][36] The paper sits at the intersection of several of her enduring concerns, including evaluation beyond multiple-choice benchmarks, the role of training data and reinforcement learning in homogenizing model outputs, and the relationship between scale and genuine reasoning ability, and was published shortly after her transition to Stanford.[35][36]
Choi's group has also explored cultural natural language processing and the cross-cultural transfer of social norms, theory-of-mind benchmarks for language models, and the use of large language model outputs as scaffolds for scientific reasoning, particularly in molecular and protein domains. These efforts are reflected in her current Stanford research portfolio and in recent invited talks at venues such as the NSF CISE Distinguished Lecture series and the Hans J. Berliner Lecture in AI at Carnegie Mellon.[1][38][39]
By 2025 her Stanford and Stanford HAI profiles described her as a co-recipient of two Test-of-Time Awards and eight Best and Outstanding Paper Awards at top AI conferences.[1][29]
Choi has become one of the most visible NLP researchers in the public conversation about AI. In April 2023 she delivered a TED main-stage talk, "Why AI is incredibly smart and shockingly stupid," which used vivid examples of large language models failing at basic commonsense reasoning to argue that scaling alone will not produce reliable intelligence, and to make the case for AI systems grounded in human norms and values.[10][11] The talk highlighted three problems with cutting-edge large language models, illustrated with examples of state-of-the-art systems failing on simple physical and social reasoning, and concluded that smaller AI systems trained on human norms and values offer a more promising path than indefinitely scaling raw web-data training.[10]
Throughout the public-facing arc of her work she has tended to frame AI capabilities and limits in terms that emphasize the gap between what large models can pattern-match and what they actually understand. In interviews tied to the MacArthur Fellowship and to Time's AI lists, she has repeatedly stressed that superior performance on one axis, such as text generation fluency, does not equate to general intelligence, and has argued for the policy importance of teaching AI systems social and moral norms specifically because of their relevance to alignment with human values.[9][26]
She has been featured in Time magazine's TIME 100 AI list in 2023 and 2025, with the 2023 profile describing her as a researcher "focused on the many ways in which human intelligence differs from that of AIs like ChatGPT" and noting that she had begun working on AI systems that understand social and moral norms specifically because of their relevance to alignment.[9] She has also been profiled in The New York Times, NPR's Short Wave, Gizmodo, and GeekWire, often in connection with Delphi and commonsense reasoning.[25][37][26][7]
Choi has delivered keynote and distinguished lectures at major venues including AAAI, ACL, cvpr, ICLR, MLSys, VLDB, WebConf, the AAMAS 2023 keynote programme, the NSF CISE Distinguished Lecture, and the Hans J. Berliner Lecture in Artificial Intelligence at Carnegie Mellon (2025).[1][38][39]
There is no reliably reported record of Choi having founded or co-founded a company outside academia; her industry experience to date consists of her tenure at nvidia Research as Senior Director of AI Research between 2024 and 2025 and her ongoing research at stanford university and Stanford HAI.[7][2][1]