Catalyzing Discovery:David, Goliath, and the future of AI
How Yejin Choi is challenging AI’s biggest assumptions—about common sense, about scale, and about who gets to build the future.
)
)
Yejin Choi. Design: Jonathan Chaves / Photo: John D. and Catherine T. MacArthur Foundation
Growing up in South Korea, Yejin Choi loved to take things apart. When something broke, she would volunteer to fix it, drawn less to the object itself than to understanding how it worked. She built model airplanes from wood and paper and remembers being the only girl to advance to a high-level city competition.
“A teacher there asked me, ‘Why are you here? Girls aren’t supposed to do this,’” she recalls. “But I wasn’t very interested in what girls were supposed to do.”
That same instinct—to resist the standard path—has defined her career. “I wanted to become a hacker,” Choi says, explaining what first drew her to computer science. Later, when she decided to leave a stable job at Microsoft to pursue a doctorate in artificial intelligence, many colleagues told her the choice made little sense.
Choi had doubts of her own. She has often described herself as an outsider, not only for her focus on unconventional problems, but because pursuing them meant entering a field—and a broader tech culture—whose norms were shaped largely without people like her in mind. This experience left an impression: “I felt like there was something wrong with me,” she says.
But over time, Choi found a way to turn that feeling into permission to keep asking the riskier questions that interested her most.
“When you have impostor syndrome so severely,” she says, “you can actually detach yourself from needing to succeed. If I’m going to fail anyway, I might as well do what I want to do.”
)
Photo: John D. and Catherine T. MacArthur Foundation
Zeroing in on common sense
A cyclist rides across a bridge suspended above nails and broken glass. Will they get a flat tire?
For a human, the answer is obvious: The danger lies below, not on the surface of the bridge. But leading AI systems have been known to answer differently, failing to grasp the structure of the situation.
For Choi—the Dieter Schwarz Foundation HAI Professor in the Department of Computer Science and the Stanford Institute for Human-Centered Artificial Intelligence (HAI)—examples like this reveal more than a technical glitch. They point to a deeper problem in how intelligence is being built. Why would a system capable of passing the bar exam fail at common sense?
The issue, she argues, is not a lack of information, but the absence of a working model for how the world behaves.
Humans interpret the world through common sense. A glass falls and shatters. A person runs, and someone else gives chase. We do not reason through these possibilities step by step; we simply know what is likely to happen next.
Modern AI, by contrast, learns from vast amounts of data to predict what comes next—and prediction is not the same as understanding. It can mimic intelligence without truly grasping how the world works.
“This is why AI models can perform some tests extremely well,” Choi says, “while surprising us with silly mistakes somewhere else.”
As Choi turned her attention to common sense, the dominant approach in AI was already coming into focus: larger datasets, more computational power, and the promise of progress through scale. Common sense did not fit neatly into that framework—it depended on context, inference, and the background knowledge humans use constantly without ever fully articulating it.
“I was a bit of a contrarian in the way I picked my research direction,” she says.
Long before today’s giant models took center stage, Choi was charting a different path through two influential projects. ATOMIC, a large commonsense knowledge graph, assembled more than a million pieces of everyday knowledge about causes, effects, intentions, and likely outcomes—an effort to make the hidden logic of ordinary life more legible to AI.
COMET built on that foundation, showing that a model trained on structured commonsense knowledge could generalize beyond what it had explicitly been given, and make plausible inferences about new situations. Together, they demonstrated that machine reasoning could improve not through brute-force scale, but through a deeper understanding of how the world works.
A different kind of Goliath
Twice named one of Time’s 100 Most Influential People in AI, Choi is one of the field’s most prominent voices—and one of its most candid critics. She describes the dominant model of AI development as a Goliath, trained on datasets with trillions of tokens, and built on the premise that bigger scale, more resources, and more brute force will eventually solve whatever problems remain.
Only a handful of companies have the power to build that Goliath. “What possibly could go wrong?” she asks wryly.
She has carried that critique well beyond the laboratory. In a September 2025 briefing to the United Nations Security Council, Choi argued that when only a few countries and companies can build the most powerful AI systems, the frontier itself becomes narrower—fewer researchers, fewer institutions, and ultimately fewer unexpected discoveries.
Her response has been to pursue concrete ways of opening that frontier. She collaborated on OpenThoughts, a large open dataset that draws researchers from multiple universities and organizations into a shared resource that smaller models can learn from.
In another line of research, she has developed techniques to distill the knowledge locked inside very large models into much smaller, more efficient ones that can reason well without requiring the vast computing resources of their giant counterparts. The aspiration, she says, is something closer to how humans actually learn.
“If you learned to drive in San Francisco,you can go to Oklahoma and start driving there right away. You don’t need to suddenly collect millions of examples of driving in Oklahoma before you turn on the car.”Yejin Choi
Smaller models use less energy and often carry a lighter infrastructure footprint, including water and land demands. They also lower the barriers to entry for researchers and institutions far beyond the handful of companies now building the largest systems.
Building a more varied AI ecosystem
But scale is not the only assumption Choi is pushing against. In work on what she calls pluralistic alignment, she and her students are exploring how AI systems might better reflect a range of human perspectives rather than collapsing toward a single “correct” answer during training.
Standard post-training methods, she argues, often smooth away that diversity, narrowing a model’s responses over time. Her research asks whether technical alternatives might preserve more of that distribution—allowing AI to better represent disagreement, coexistence, and the complexity of human values rather than flattening them into one.
Looking ahead, Choi is applying these ideas to scientific research—a domain where the conditions could hardly be more different from the data-rich environment that produced today’s giant models. Many scientific fields are defined by scarcity: limited data, complex structures, and large gaps in what is known.
In chemistry, for instance, a model might reason backward from a target compound and propose a synthesis pathway for a molecule that has never been made before. That is where the promise of AI in science becomes most consequential: not in producing better summaries of the known world, but in helping researchers imagine—and test—what lies beyond it.
Taken together, these projects point toward a broader vision for AI: not a single dominant system, but a more varied ecosystem of models—small, medium, large, each useful in different ways and accessible to different communities.
That vision has found a natural home at Stanford HAI, where the question is not only how powerful AI can become, but how it can better serve human needs. It also echoes something older in Choi herself—a lifelong skepticism of systems that insist there is only one right way to build, think, or belong.
Human-Centered Artificial Intelligence: The impact
Why it matters
The most powerful AI systems in the world are built by a handful of companies in a handful of countries. Choi’s research makes the case for something different: AI that is smaller, smarter, less resource-intensive, and open to far more people.
The opportunity
Yejin Choi is advancing a more human-centered vision for AI—one that challenges the assumption that bigger is always better. Support for Stanford HAI helps make possible the kind of bold, interdisciplinary research Choi is pursuing—from commonsense reasoning and small language models to scientific applications that could make AI more accessible, more trustworthy, and more responsive to human needs.
)
)
:focal(745x365:746x366))
:focal(740x1051:741x1052)){kind=spacer}