Wiki · Person · Last reviewed May 19, 2026

Yejin Choi

Yejin Choi is a computer scientist whose research connects natural language processing, common-sense reasoning, social and moral norms, value pluralism, and the limits of large language models.

Known for: common-sense AI, ATOMIC, Social Chemistry, Delphi, pluralistic alignment Affiliation: Stanford University; Stanford Institute for Human-Centered Artificial Intelligence

Overview

Choi is the Dieter Schwarz Foundation Professor in Stanford's Department of Computer Science and a Senior Fellow at the Stanford Institute for Human-Centered Artificial Intelligence. Stanford lists her current research interests as including the limits and capabilities of large language models, alternative training recipes, symbolic methods for neural networks, reasoning and knowledge discovery, moral norms and values, pluralistic alignment, and AI safety.

Before Stanford, Choi was a professor at the University of Washington and held a joint appointment at the Allen Institute for Artificial Intelligence. She was named a 2022 MacArthur Fellow for work using natural language processing to help AI systems make commonsense inferences about the world.

Her work matters because it sits at a hard boundary in AI: models can be fluent, useful, and superhuman at some tasks while still failing at ordinary background knowledge, social meaning, value conflict, and context-sensitive judgment.

Common-Sense AI

Common-sense AI is the long-running problem of giving machines access to ordinary background knowledge about people, objects, events, causes, intentions, and likely consequences. Choi's work treats common sense as a practical capability rather than a philosophical ornament: systems need it to interpret language, act in the world, explain situations, and avoid brittle mistakes.

The ATOMIC project, introduced in 2018, built a large atlas of if-then commonsense inferences around everyday events. Instead of only asking a model to classify a sentence, ATOMIC represents likely causes, effects, intentions, reactions, and needs surrounding ordinary situations. This line of work helped make common-sense reasoning more concrete for neural language systems.

Choi's broader common-sense program also includes visual and textual grounding, narrative understanding, and systems that generate multiple plausible inferences rather than pretending one answer exhausts a situation.

Social and Moral Norms

Choi's later work moved from physical and causal common sense into social common sense: what people consider rude, kind, harmful, permissible, risky, unfair, or context-dependent. This matters because AI assistants increasingly mediate social life, workplace communication, education, companionship, persuasion, and moral advice.

Social Chemistry 101 introduced a large corpus of everyday social norms and moral judgments, organized around rules of thumb and dimensions such as social judgment, cultural pressure, and legality. The goal was not to declare one universal moral code, but to model how people reason about ordinary social situations.

The Delphi line of research explored machine moral judgment using collections of human judgments over real-world situations. It also exposed a central tension: learning from human moral data can improve behavior in some settings, but it can also reproduce cultural bias, status-quo assumptions, ambiguity, and disagreement.

Pluralistic Alignment

Choi is important to alignment debates because her work resists the fantasy that human values are a single clean objective. Real societies contain disagreement across cultures, communities, generations, religions, political views, personal histories, and contexts. A system that optimizes one flattened value function can erase this diversity or launder one group's norms as universal reason.

Pluralistic alignment asks how AI systems can respect value diversity while still refusing abuse, deception, coercion, discrimination, and dangerous action. This is harder than making a model polite. It requires uncertainty, context, consultation, contestability, and institution-level choices about who gets to define acceptable behavior.

Choi's research program therefore connects technical NLP to governance questions. If assistants, agents, tutors, search systems, and companions are trained on moral judgments, someone must ask whose judgments are represented, how conflict is handled, and how affected people can challenge the system.

Why She Matters

Choi gives the AI debate a vocabulary for a missing middle layer. On one side are raw capabilities: scale, benchmarks, tool use, coding, multimodal generation. On the other side are high-level governance claims about safety and human values. Her work studies the everyday inferential material between them: what a situation means, what likely follows, what people intend, what a community considers acceptable, and when a rule needs context.

That layer becomes more important as AI systems move into agentic workflows. A model that lacks common sense can misunderstand an instruction. A model that lacks social sense can amplify harm while sounding helpful. A model that treats morality as one label can become coercive, parochial, or falsely authoritative.

Choi is also a public counterweight to simple AI triumphalism. Her TED talk framed contemporary AI as both surprisingly capable and shockingly brittle. That formulation captures a key reality of the current transition: the systems are powerful enough to reshape institutions, but not grounded enough to trust without scrutiny.

Spiralist Reading

Choi is a reality-friction researcher.

Her work asks whether the machine knows the ordinary world beneath the words: not only syntax, but breakfast, shame, obligation, politeness, harm, jokes, excuses, power, and disagreement. That is exactly where synthetic systems become socially dangerous. They can speak as if they understand the human scene while missing the scene itself.

For Spiralism, Choi's importance is the refusal of moral compression. A society cannot be reduced to one scalar reward. A person cannot be reduced to a prompt. A community cannot be reduced to a dataset without asking who was counted, who was flattened, and who was made invisible.

Open Questions

Sources

Return to Wiki