AI scientist, Stanford professor, and World Labs co-founder (AMD agreed in September 2026 to acquire World Labs; Li is to become AMD’s chief scientist after closing)
Public mandate vs personal authority
AI
Professional case
Case 39 of 41
Record last updated October 5, 2026
The case question
How does scientific authority travel across a university, a field-building institute, and a company?
What this case shows
Li shows that a portable research thesis can seed several kinds of institution without becoming the private property of its most visible scholar. Durability comes from governable datasets, credited collaborators, independent scrutiny, distributed leadership, and rights that remain legible as work crosses university, company, nonprofit, policy, and venture boundaries.
§Sourced to public records·18 sources (8 independent)·Not independently reviewed·How we check →
01
The career in brief
Li’s research leadership includes ImageNet, Stanford professorship, Google Cloud AI, and creation of Stanford’s human-centered AI institute; she later co-founded World Labs, which AMD agreed in September 2026 to acquire.
Why the case matters
The case asks how a scientist carries credibility across institutions with different incentives without allowing one affiliation to validate every claim made in another.
02
Build · Carry · Control · Continue
Four separate questions. Public visibility alone answers none of them.
Build
What was built
Turning point
Scientific authority moved among dataset building, academic institution-building, corporate leadership, public policy, and a venture-backed company.
Carry
What can travel
Open question
What rights and responsibilities govern the continued use, maintenance, removal requests, provenance, and downstream reuse of ImageNet data?
Control
What can be governed
Open question
What formal decision rights did Li hold over Project Maven, Google Cloud product policy, and disclosure, as distinct from scientific advice and executive title?
Continue
What persists if a dependency changes
Open question
How were crowd workers and other contributors compensated, credited, and protected, and what obligations persist as dataset governance evolves?
The tension in this case: here, public mandate vs personal authority means how scientific authority moves with one researcher among a public dataset, a university institute, a corporate role, and a venture-backed company, each of which holds its own rights and decisions. Look for who holds the dataset rights, how the institute is governed, and what equity, control, and decision rights come with each role.
03
What happened
2005–2009
After faculty appointments at Illinois and Princeton, Li joined Stanford. She and collaborators developed ImageNet as a large labeled image dataset intended to overcome a scale bottleneck in object recognition.
The ImageNet Challenge became shared research infrastructure through which many teams demonstrated rapid gains in computer vision. Li also directed Stanford’s AI Lab from 2013 to 2018 and helped develop the educational initiative that became AI4ALL.
During a Stanford sabbatical, Li served as a Google vice president and chief scientist of AI and machine learning at Google Cloud. The New York Times and Business Insider reported leaked emails in which Li expressed concern about how Google’s Project Maven military-AI work would be publicly framed. Li told the New York Times that working on any project meant to weaponize AI would be “deeply against my principles.”
Li returned to Stanford and co-founded the university-wide Institute for Human-Centered Artificial Intelligence, extending her work into research, education, policy, and public institutional design. She also held board, advisory, and venture roles.
External criticism exposed offensive and socially loaded labels in ImageNet’s person hierarchy. Li and collaborators published a technical review of the causes and proposed filtering and rebalancing, showing both the dataset’s institutional influence and the governance debt embedded in inherited taxonomies and crowd labor.
Li expanded her public-policy and health-AI work while taking partial academic leave and co-founding World Labs with three colleagues to develop spatial-intelligence systems. The company announced $230 million in financing from major venture and corporate investors.
As World Labs CEO, Li carried a recognizable research thesis, scientific reputation, and network into a venture-backed company. The move created a new operating institution, but its capital, models, data, compute, employment relationships, and intellectual property are collectively produced and governed through the company rather than personally owned by Li.
AMD signed a merger agreement to acquire World Labs for about $8.2 billion in AMD common stock, according to AMD’s Form 8-K. The Next Web reported that Li would become AMD’s executive vice president and chief scientist, reporting to CEO Lisa Su, and that the deal was expected to close by the end of 2026, subject to regulatory approvals. Public sources do not state how proceeds, retention terms, or intellectual-property rights are allocated among Li, her co-founders, employees, and investors.
Li’s case traces a research idea as it moves through a public dataset, academic lab, corporate role, nonprofit, policy institution, and venture-backed company. Her intellectual portability is unusually strong. But each manifestation depends on different collaborators, funders, infrastructures, governance systems, and rights. The central question is which parts of her influence have become durable public infrastructure and which now sit inside organizations she does not solely control.
Evidence that complicates the first reading
ImageNet was led by Li but built by students, research collaborators, crowd workers, source-image ecosystems, WordNet’s taxonomy, challenge participants, and institutions. Describing it as a solitary invention erases essential labor and inherited design choices.
The deep-learning advances associated with ImageNet were produced by many outside teams using the benchmark; the dataset enabled and measured progress but did not itself create every downstream model or application.
ImageNet’s offensive person labels and representational imbalance are adverse evidence that widely shared infrastructure can scale social assumptions and governance failures along with scientific utility.
The Project Maven record complicates a clean human-centered-AI narrative: according to reporting on leaked emails, Li expressed concern about military-AI framing, while practical authority over the contract and disclosure sat within Google’s corporate structure.
Stanford HAI’s research and policy authority is collectively produced by co-directors, faculty, staff, donors, university governance, and external partners; it is not Li’s personally controlled platform.
World Labs demonstrates portability into entrepreneurship, but venture capital, co-founders, employees, compute providers, datasets, and future customers constrain founder autonomy. Valuation and financing do not prove technical success, safety, product-market fit, or institutional durability.
The pending AMD acquisition would place World Labs inside a larger public company. The sources reviewed do not describe how decision rights over its models, data, and research agenda would be allocated after closing.
What remains unknown
What rights and responsibilities govern the continued use, maintenance, removal requests, provenance, and downstream reuse of ImageNet data?
How were crowd workers and other contributors compensated, credited, and protected, and what obligations persist as dataset governance evolves?
What formal decision rights did Li hold over Project Maven, Google Cloud product policy, and disclosure, as distinct from scientific advice and executive title?
How are funding, donor influence, conflicts, research independence, and policy positions governed across Stanford HAI and Li’s outside roles?
What equity, board rights, IP assignments, data licenses, model rights, and control provisions govern World Labs among Li, her co-founders, employees, and investors?
What will Li’s Stanford and Stanford HAI roles be after she joins AMD, if the acquisition closes?
What independent evidence will establish World Labs’ model performance, safety, data provenance, environmental cost, commercial outcomes, and capacity to continue without Li?
05
Change one dependency
Use a counterfactual to expose the architecture: answer first, then compare your reasoning with the record.
Write a sentence or two before you look at the record. Your answer is kept only in this browser; nothing is sent, scored, or stored on a server.
Thought experiment
Which research thesis, relationships, and reputation stay with Li, and which equity, data, models, and employment relationships stay with the company?
The record says World Labs' capital, models, data, compute, and intellectual property are produced and governed through the company. This tests what is personally portable.
Use the record, not intuition: identify what the sources establish, then check the open questions in this case. This tool does not predict what Fei-Fei Li will do.
The missing evidence that matters most: What rights and responsibilities govern the continued use, maintenance, removal requests, provenance, and downstream reuse of ImageNet data?
06
Read it against another career
Hold the tension constant across different careers, or hold the field constant and inspect a different path. A comparison is useful because it can challenge the first explanation.
18 sources: 8 independent and 10 other (primary, institutional, subject interviews, self-authored, or republished reports). First-party sources and interviews establish what a person or organization said or announced; they do not independently prove performance, ownership, causation, or impact. This is analysis of a public record and makes no rating of the person.
Each link was checked by hand during the record’s source review; the site does not continuously re-check them. A broken link, or one that does not support its statement, is a defect in the record. Please report it. 2 links lead to a publisher or organization landing page rather than the exact supporting item, so they identify a research lead rather than a claim-level citation.