Fei-Fei Li

Fei-Fei Li (born 1976) is a Chinese-American computer scientist and the Sequoia Professor at Stanford University, where she co-directs the Institute for Human-Centered Artificial Intelligence. Her most influential contribution is ImageNet, a labelled database of millions of images that made large-scale visual recognition possible and helped trigger the deep-learning revolution 1.

Li's career spans the technical and the institutional. She built the dataset that others used to demonstrate the power of neural networks, spent a year and a half as chief scientist for AI at Google Cloud, and has become a public advocate for an approach she calls human-centred AI — one that treats the technology as a tool for augmenting people rather than replacing them 2.

In 2024 she co-founded World Labs, a company working on what she calls spatial intelligence: systems that understand and generate three-dimensional environments. This entry covers her education, ImageNet, the deep-learning turn, her work at Google and Stanford, World Labs, her stated views, and the criticism her record has attracted.

Li was born in Beijing in 1976 and grew up in Chengdu. At fifteen she moved to the United States, settling in Parsippany, New Jersey, and attended high school there while learning English. She has written about the adjustment in her memoir, describing a period of economic precarity and cultural dislocation that shaped her later interest in making technology serve people outside its usual centres of power 3.

She studied physics at Princeton University, graduating in 1999, and then moved to the California Institute of Technology, where she earned a doctorate in electrical engineering in 2005. Her graduate work combined physics, neuroscience and computer vision, an interdisciplinary range that remained characteristic of her later research 1.

She joined the faculty at Princeton in 2005 and moved to Stanford in 2009, where she has remained. Her early work was on visual recognition, scene understanding and the statistical structure of images, at a time when computer vision relied on hand-engineered features and classifiers trained on small datasets 3.

She has said that the question that drove her was not how to build a better classifier but how to give machines the kind of visual experience a child accumulates. That framing led directly to the dataset work for which she is best known 4.

In 2006 Li began work on a dataset that would cover the visual world at a scale no previous benchmark had attempted. The project used WordNet, a lexical database in which nouns are organised in a hierarchy, to define tens of thousands of object categories, then collected and labelled images for each one. The result, published in 2009, contained more than three million images across thousands of categories 5.

The key insight was that diversity mattered as much as size. Earlier datasets were small and often centred on a few classes; ImageNet covered the long tail of objects, in many poses, settings and lighting conditions. That variety was what allowed a model trained on it to generalise, and it was the reason the dataset became the standard benchmark for visual recognition 6.

Labeling was the hard part. Li's team used Amazon Mechanical Turk to distribute the annotation work, building quality controls that could catch error and inconsistency. The process was expensive and slow by the standards of later synthetic data, but it produced a resource that researchers used for more than a decade 5.

The ImageNet Large Scale Visual Recognition Challenge, run from 2010, turned the dataset into a competition. In 2012 AlexNet, a convolutional network trained on the data, cut the top-five error rate dramatically, and the result is widely treated as the beginning of the modern deep-learning era 6. Li has consistently described ImageNet as a collective achievement rather than a solo one. The dataset was built by a team, the challenge was run with collaborators, and the breakthrough it enabled came from other researchers. Her contribution was to recognise that data scale was the bottleneck and to spend years removing it 4.

After 2012 Li's research shifted toward the new methods. Her group worked on image captioning, visual question answering and the combination of vision with language, areas that became foundational for later multimodal systems. The work was part of a broader move in computer vision from hand-designed features to learned representations 1.

She also became interested in the human side of the technology. As visual systems moved into surveillance, hiring and medical imaging, the questions of bias, consent and accountability became harder to treat as someone else's problem. That shift in emphasis, from making vision work to asking what it should be used for, defined the second half of her career 2.

In 2017 she took leave from Stanford to serve as chief scientist for artificial intelligence and machine learning at Google Cloud. The role put her inside a company whose AI contracts with the US military were becoming controversial among its employees, and she was involved in internal discussions about how the company should engage with defence work 3.

She returned to Stanford in 2018 and co-founded the Institute for Human-Centered Artificial Intelligence, which launched in 2019 with the aim of bringing together computer science, law, medicine, economics and the humanities. The institute's premise is that AI is a societal technology, not a narrow engineering discipline 2.

HAI was founded with an unusual structure for a university centre: it funds research across schools, runs policy briefings, and publishes an annual AI Index that tracks technical progress, investment, and social impact. Under Li's co-direction it has become one of the most visible academic voices in AI policy, particularly in the United States 2.

The institute's stated position is that AI should augment human capability rather than replace it. That sounds uncontroversial, but it implies a research agenda: systems designed to assist clinicians, teachers and scientists, evaluation that measures human outcomes rather than benchmark scores, and governance that includes the people affected by deployment 4.

Li has testified before Congress and advised governments, arguing for public investment in AI research and for a governance approach that protects civil rights without freezing the technology. She has been critical of both unrestricted deployment and blanket moratoria, and has described the choice between innovation and safety as a false dichotomy 3.

Her memoir, published in 2023, presented the argument in personal terms. It described her family's immigration, her early research and her conviction that the people who build AI should reflect the diversity of the people who use it. The book was widely reviewed and raised her public profile beyond the research community 4.

In 2024 Li co-founded World Labs, a company developing what it calls large world models: systems that reconstruct, generate and reason about three-dimensional scenes. The premise is that current AI is strong with language and flat images but weak with the spatial structure of the physical world, and that closing that gap is necessary for robotics, design and scientific simulation 7.

The company emerged from research on neural radiance fields and Gaussian splatting, techniques that reconstruct a scene from photographs. World Labs' contribution, as described publicly, is a model that generates persistent, explorable 3D environments from an image or a description, with consistent geometry rather than a video-like illusion 7.

Li has framed spatial intelligence as the next step after language models, not as a replacement for them. Her argument is that understanding the world requires more than predicting text, and that the ability to represent objects, space and physical relationships is a prerequisite for embodied systems and for scientific applications 4.

The company raised venture funding in 2024 and 2025 and has not yet shipped a commercial product. Its significance at this stage is as a statement of research direction: one of the most prominent academic vision researchers has concluded that the next frontier is three-dimensional understanding, and has left a comfortable position to pursue it 7.

Li's public positions centre on the phrase "human-centred AI". She argues that the technology should be evaluated by what it does for people — in health, education, accessibility and scientific discovery — rather than by benchmark scores or by the abstract goal of matching human intelligence. She has said that there is nothing artificial about AI, because it is made by people and affects people 2.

She is sceptical of both utopian and apocalyptic framings. On her account, the near-term risks are concrete: bias in medical algorithms, surveillance without consent, concentration of power in a few companies, and a workforce that does not reflect the population. She has criticised the AI safety debate for focusing on speculative scenarios while neglecting harms that are already documented 4.

She has argued for public investment in AI research and for universities to retain a role in fundamental work, warning that the field's centre of gravity has moved to industry and that the questions industry does not ask may go unstudied. She has also called for greater diversity in the teams that build the systems, not as a matter of representation alone but because homogeneous teams produce systems that fail for everyone else 3.

On governance, she favours sector-specific regulation, transparency about training data and evaluation, and international cooperation on standards. She has been careful not to endorse specific legislative proposals, which has disappointed some advocates and protected her ability to work with policymakers across positions 2.

ImageNet has attracted sustained criticism. Researchers, including Timnit Gebru and others, documented that the dataset contains derogatory labels, biased categories and images collected without consent, and that models trained on it inherit those problems. Li has acknowledged the issues, supported work on dataset documentation, and argued that the dataset should be understood as a product of its time rather than as a neutral resource 3.

The critique is not only about labels. The dataset's categories reflect a particular view of the world, and its images were scraped from the web, raising questions about privacy and copyright that are still unresolved. Defenders note that ImageNet was built before those norms existed and that its documentation has been revised; critics reply that the norms were not unknown at the time 6.

Her time at Google coincided with employee protests over the company's military contracts, and she was criticised by some employees for her handling of the internal debate. She has said that she supported transparency and that the episode illustrated the difficulty of aligning corporate work with ethical commitments. The record of what was said internally is contested, and she has not discussed it in detail publicly 3.

Her "human-centred" framing has been criticised as vague. The phrase can accommodate positions from regulation to deregulation, and some researchers argue that it functions as a brand rather than a research programme. HAI's industry funding has also drawn questions about independence. Li responds that the institute's funding is disclosed and that engagement with industry is necessary if the research is to matter 2.

How much credit ImageNet deserves is debated. One account treats the dataset as the pivotal contribution that made deep learning possible; another notes that convolutional networks, GPU training and the backpropagation algorithm were developed independently, and that ImageNet was the testbed rather than the cause. Li's own account is closer to the second: she describes the dataset as removing a bottleneck, not as inventing the method 5.

The assessment of her policy role is likewise split. Supporters describe her as a bridge between technical research and public institutions, someone who has made AI legible to legislators without exaggerating either the benefits or the risks. Critics describe the bridge as too comfortable with power, noting that HAI takes industry money and that its recommendations rarely threaten large companies 2.

The Google episode is the least settled. Public reporting describes internal debate about military work and about the company's public position; Li has spoken about the importance of engagement and about the difficulty of the choices, but the full record has not been published. What is clear is that she left the role after a short tenure and returned to academia, and that the controversy followed her 3.

What is not disputed is the effect of her work on the field's direction. ImageNet changed what a benchmark could be, demonstrated the value of large labelled datasets, and gave the deep-learning community a common target. Even the criticisms of the dataset are a measure of its influence: nobody writes papers about the politics of a dataset that no one uses 6.