← AI Safety Orgs

Stanford HAI

Research

Human-centered AI framing.

Founded
2019
HQ
Stanford, CA
Structure
university-affiliated
Model
Mixed

Theory of Change

Stanford HAI believes AI risk is best addressed through interdisciplinary academic research, informed governance, and transparency measurement. In Li's words: "AI is a tool, and its values are human values." The causal chain: (1) produce authoritative public measurements (AI Index, FMTI, HELM) that inform policymakers, (2) train legislators directly through congressional boot camps and expert testimony, (3) level the compute playing field between industry and academia via NAIRR, and (4) fund cross-disciplinary research that incorporates social science, ethics, and policy perspectives into AI development.

Li explicitly deprioritizes existential risk relative to near-term harms. In a 2023 MIT Tech Review interview: "I absolutely respect that [x-risk]. But if you ask me as an AI leader... I feel there are other risks that are what I would call catastrophic risks to society that are more pressing and urgent." She highlights misinformation, workforce disruption, bias, and privacy. At the 2025 Paris AI Summit, she said governance should be based on "science, not science fiction."

However, in March 2025 Li co-led a California policy working group that recommended AI laws should "anticipate future risks" including those "not yet been observed in the world," stating: "If those who speculate about the most extreme risks are right -- and we are uncertain if they will be -- then the stakes and costs for inaction on frontier AI at this current moment are extremely high." This represents a notable shift from her SB 1047 opposition seven months earlier.

What They Do

Measurement and transparency tools:

  • AI Index Report (8 annual editions, 2018-2025): the most widely cited comprehensive survey of AI progress, used as a policy reference globally.
  • Foundation Model Transparency Index (FMTI): scores 13 major AI companies on 100-point transparency scale. The 2025 edition found transparency DECLINED from 58 to 40 average. Meta dropped from 60 to 31; IBM scored 95; xAI and Midjourney scored 14.
  • HELM/HELM Safety: holistic evaluation framework for language models, including safety benchmarks across 6 risk categories.

Policy engagement:

  • NAIRR: HAI's flagship policy achievement. Li and Etchemendy proposed the concept in 2019. HAI organized 22 universities, published a blueprint, served on task forces, and shepherded it into legislation (CREATE AI Act, 2023) with an NSF pilot launched.
  • Congressional Boot Camp: 76+ staffers trained in 3-day intensive programs since 2022.
  • Tech Ethics & Policy Fellowship: places Stanford students in DC offices.
  • Li and Senior Fellow Rob Reich met with President Biden on AI policy.

Research funding distributed:

  • $50M+ to 400+ Stanford scholars across all seven schools.
  • $27.6M via Hoffman-Yee Research Grants (up to $500K/team year 1).
  • $10M in industry-sponsored research grants; $9M in cloud credits.

SB 1047 opposition: Li published a Fortune op-ed (Aug 2024) arguing the bill would "harm our budding AI ecosystem" and "shackle open-source development." Senator Wiener and others identified specific inaccuracies in her arguments. Bengio wrote a direct rebuttal. After the bill's veto, Li co-led the governor's working group that recommended transparency and testing requirements echoing much of what SB 1047 would have accomplished.

CRFM (Center for Research on Foundation Models, dir. Percy Liang): HAI's most safety-relevant sub-center. Authored the 212-page "Opportunities and Risks of Foundation Models" report (2021). Runs HELM and FMTI. Liang receives Open Philanthropy funding for alignment and safety evaluation research.

Key People

Fei-Fei Li -- Co-Director (founding). Created ImageNet. Princeton BA (physics), Caltech PhD. VP/Chief Scientist at Google Cloud 2017-2018, where leaked emails showed she privately praised Project Maven but warned against mentioning AI for PR reasons. Co-founded World Labs in 2024 (spatial intelligence AI, raised $1B+, valued approaching $5B). On partial academic leave from Stanford 2024-2025. Her personal story -- Chinese immigrant, worked in family dry cleaning shop through Princeton -- is genuinely compelling and shapes her human-centered philosophy.

Percy Liang -- Senior Fellow, Director of CRFM. Leads HELM, FMTI, and alignment research. Co-founded Together AI (open-source AI platform). Largest individual recipient of Open Philanthropy AI safety grants at Stanford (~$6.1M). HAI's strongest connection to the AI safety ecosystem.

Russell Wald -- Executive Director (promoted from Deputy Director). Key operational leader and NAIRR champion. Appears to run day-to-day operations while Li splits time with World Labs.

Advisory Council chaired by Reid Hoffman (LinkedIn co-founder, Microsoft board), with Condoleezza Rice, Eric Horvitz (CSO Microsoft), James Manyika (SVP Google/Alphabet), John Hennessy (Alphabet board chair). ~35 people on faculty and staff; 22+ senior fellows.

Money and Incentives

Total budget: Unknown. HAI is part of Stanford University with no separate financial filings. This is the most significant structural finding: an institute that grades AI companies on transparency does not publish its own budget.

Known revenue streams:

  • Corporate Founding Members: $5M/year each, 3-year commitments (Google, IBM confirmed; full list not public)
  • Corporate Affiliate Program: $550-800K/year each (Accenture, McKinsey, AXA, LVMH, SCBX, Infosys confirmed)
  • Hoffman-Yee philanthropic grants: $27.6M cumulative, likely ~$5M/year
  • Federal research grants: amount unknown
  • Individual philanthropy: amount unknown
  • Cloud credits from Google ($100K/researcher) and Microsoft Azure ($50K/researcher): $9M total delivered

Rough estimate: $25-35M/year operating budget, though this could be significantly higher.

Incentive analysis:

  • Corporate funding is likely HAI's single largest revenue source. The advisory council includes Microsoft CSO, Google SVP, and Alphabet board chair -- the same companies providing compute credits and subject to FMTI scoring.
  • Li's World Labs (approaching $5B valuation) creates a direct financial interest in AI capabilities growth. Liang's Together AI creates similar tensions with his safety evaluation work.
  • Stanford's endowment ($60.1B) has ~$21.6B in private equity, heavily AI-exposed. OpenAI's CFO recently joined Stanford's Board of Trustees. The university itself has structural incentives against positions that threaten AI valuations.
  • The Project Maven emails are the clearest evidence of how these incentives operate in practice: Li's concern about the military AI contract was reputational damage to Google Cloud's "Humanistic AI" brand, not the ethics of drone targeting.
  • HAI's SB 1047 opposition aligned with every major corporate partner's position.

Open Philanthropy funding ($9.3M to Stanford AI-relevant): Goes to individual researchers (primarily Liang, Hashimoto, Potts, Barrett), not to HAI as an institution. This creates an interesting dual funding structure: HAI's most safety-relevant researchers are funded by the EA/safety ecosystem while the institution is funded by industry.

What Others Say

Yoshua Bengio (Turing Award, direct rebuttal to Li): "I disagree with her recently published stance on SB 1047... We cannot let corporations grade their own homework and simply put out nice-sounding assurances. We don't accept this in other technologies such as pharmaceuticals, aerospace, and food safety."

Gary Marcus (open letter to Li): Li "favors AI governance, but doesn't make any positive, concrete suggestion for how to address risks such as mass casualties, weapons of mass destruction, large-scale cyberattacks."

Nathan Lambert (Interconnects/EleutherAI, FMTI critique): FMTI "measures how well-documented a commercial product is" rather than true transparency. "If this Index were adopted as regulation, it would be a textbook example of regulatory capture."

Margaret Mitchell (at HAI's own workshop): Scrolled through HAI's collaborator roster showing homogeneity. Argued the "foundation models" naming was a "rebrand" serving industry. These models are "support structures, not foundations."

Senator Scott Wiener (on Li's SB 1047 claims): "It was crystal clear in the bill that you're only required to shut down a model if it is in your possession. And yet, Fei-Fei Li put that inaccurate statement in her piece. She's very well respected, so it was unfortunate."

Ruth substack (on Stanford's Gebru silence): "Was Stanford's silence on Timnit Gebru the price of access to industry compute power?"

Stanford Daily (internal): Recommended Stanford stress-test its AI financial exposure and "keep strong firewalls between corporate interests and academic decisions."

What's Absent

  • No public HAI budget or financial transparency despite running a transparency index for AI companies.
  • No explicit research agenda or strategic plan addressing existential or catastrophic AI risk.
  • Complete list of corporate founding members not public.
  • No conflict-of-interest disclosures for Li's World Labs or Liang's Together AI.
  • No HAI leaders appear on 80,000 Hours, Dwarkesh Patel, or other EA/safety podcasts -- zero engagement with safety community media.
  • No HAI position on CAIS extinction risk statement, pause proposals, or any x-risk-specific framework.
  • No published evaluation of whether HAI's interventions (NAIRR, FMTI, bootcamps) have changed outcomes.
  • No notable departures with public criticism of HAI's direction.
  • FMTI's own results show transparency declining -- measurement alone is not changing behavior.

Recommended Reading

  1. Leaked Emails: Google Project Maven (The Intercept, May 2018) -- The single most revealing document about the gap between HAI leadership's public and private positions on industry-academy conflicts. Li praised the military AI contract internally while warning about PR damage; publicly claimed it violated her principles. https://theintercept.com/2018/05/31/google-leaked-emails-drone-ai-pentagon-lucrative/

  2. Bengio rebuttal of Li on SB 1047 (Fortune, Aug 2024) -- The strongest direct counterargument to HAI's regulatory position, from a peer of equal stature. https://fortune.com/2024/08/15/yoshua-bengio-californias-ai-safety-bill-will-protect-consumers-innovation-tech/

  3. FMTI critique by Nathan Lambert et al. (Interconnects, Oct 2023) -- Substantive 5,000-word critique arguing FMTI measures documentation compliance rather than true transparency and could facilitate regulatory capture. https://www.interconnects.ai/p/fmti-critique

  4. Fei-Fei Li on Tim Ferriss (Dec 2025) -- 11K-word transcript. The most personal and candid Li interview. No safety questions asked, but reveals the worldview and motivations of the person running HAI. https://tim.blog/2025/12/10/dr-fei-fei-li-the-godmother-of-ai-transcript/

  5. FMTI 2025: Transparency on the Decline (HAI, Dec 2025) -- HAI's own finding that transparency has declined despite measurement. Raises the question of whether their theory of change works. https://hai.stanford.edu/news/transparency-in-ai-is-on-the-decline

Connected to (10)

EleutherAIcollaboratorOpen Philanthropycollaborator · Percy Liang
World Labsstaff to · Fei-Fei Li
Together AIstaff to · Percy Liang
Alphabetboard overlap · John Hennessy
Googlecompute provider
Microsoftcompute provider · Eric Horvitz
Stanford Center for AI Safetycollaborator · Clark Barrett
Google Cloudstaff from · Fei-Fei Li
AI4ALLspun off from · Fei-Fei Li
Sources (71)
Every URL that was read during research.
  1. 1.Abouthai.stanford.edu
  2. 2.Researchhai.stanford.edu
  3. 3.Industryhai.stanford.edu
  4. 4.The 2025 AI Index Reporthai.stanford.edu
  5. 5.Stanford HAI at Five: Pioneering the Future of Human-Centered AIhai.stanford.edu
  6. 6.At five years old, Institute for Human-Centered AI looks to the futurestanforddaily.com
  7. 7.Fei-Fei Li - Wikipediaen.wikipedia.org
  8. 8.HELM Safety: Towards Standardized Safety Evaluations of Language Modelscrfm.stanford.edu
  9. 9.National AI Research Resource (NAIRR)hai.stanford.edu
  10. 10.Introducing New Faculty, Staff Leaders at Stanford HAIhai.stanford.edu
  11. 11.“AI Is a Tool, and Its Values Are Human Values.”issues.org
  12. 12.People of HAIhai.stanford.edu
  13. 13.Stewarding AI in the Age of Big Techhai.stanford.edu
  14. 14.Behind the $1.98 billion in Stanford's external research fundsstanforddaily.com
  15. 15.How Unique Are the Risks Posed by Artificial Intelligence?hai.stanford.edu
  16. 16.On Independence and Influencehai.stanford.edu
  17. 17.The Tim Ferriss Show Transcripts: Dr. Fei-Fei Li, The Godmother of AI — Asking Audacious Questions, Civilizational Technology, and Finding Your North Star (#839) - The Blog of Author Tim Ferrisstim.blog
  18. 18.Percy Liang, Stanford: On the paradigm shift and societal effects of foundation modelsimbue.com
  19. 19.Transparent and Interpretable AI: an interview with Percy Liang - Future of Life Institutefutureoflife.org
  20. 20.AI Godmother Fei-Fei Li Has a Vision for Computer Visionspectrum.ieee.org
  21. 21.Shaping AI Benchmarks with Together AI Co-Founder Percy Liangwandb.ai
  22. 22.How the Foundation Model Transparency Index Distorts Transparencyinterconnects.ai
  23. 23.Percy Liang on truly open AIpress.airstreet.com
  24. 24.On the Opportunities and Risks of Foundation Modelscrfm.stanford.edu
  25. 25.SAIAstanfordaialignment.org
  26. 26.Stanford Center for AI Safetyaisafety.stanford.edu
  27. 27.The Tech Coup: A New Book Shows How the Unchecked Power of Companies Is Destabilizing Governancehai.stanford.edu
  28. 28.Is a “Rogue AI” Catastrophe Coming? Stanford’s Mark Lemley on California’s AI Safeguards Bill | Stanford Law Schoollaw.stanford.edu
  29. 29.All Eyes on Sacramento: SB 1047 and the AI Safety Debatecarnegieendowment.org
  30. 30.Policyhai.stanford.edu
  31. 31.An open letter to Fei-Fei Li concerning California’s proposed AI regulation, SB-1047garymarcus.substack.com
  32. 32.Why Silicon Valley is trying so hard to kill this AI bill in Californiacalmatters.org
  33. 33.Stanford Research Teams Receive New Hoffman-Yee Grant Funding for 2025hai.stanford.edu
  34. 34.Grantshai.stanford.edu
  35. 35.Azure Cloud Credit Grantshai.stanford.edu
  36. 36.Google Cloud Credit Grantshai.stanford.edu
  37. 37.Stanford AI Scholars Find Support for Innovation in a Time of Uncertaintyhai.stanford.edu
  38. 38.Stanford Team Develops a Blueprint for a National AI Research Resourcehai.stanford.edu
  39. 39.Decoding the White House AI Executive Order’s Achievementshai.stanford.edu
  40. 40.Stanford HAI Welcomes Six Distinguished Scholars as Senior Fellowshai.stanford.edu
  41. 41.AI, Ethics, and Public Service: Bridging the Gap through Stanford HAI’s Fellowship Programhai.stanford.edu
  42. 42.Google's AI Cloud Star Leaves After Pentagon Deal Protestsdatacenterknowledge.com
  43. 43.AI keeps getting more powerful, making it harder to judge how smart models actually arefortune.com
  44. 44.Expanding Academia’s Role in Public Sector AIhai.stanford.edu
  45. 45.Transparency in AI is on the Declinehai.stanford.edu
  46. 46.Progress Toward Safe and Reliable AIhai.stanford.edu
  47. 47.Unknownaisafety.stanford.edu
  48. 48.Unknownhai.stanford.edu
  49. 49.AI is at an inflection point, Fei-Fei Li saystechnologyreview.com
  50. 50.Leaked Emails Show Google Expected Lucrative Military Drone AI Work to Grow Exponentiallytheintercept.com
  51. 51.Geoffrey Hinton and Fei-Fei Li in conversationdigitalhabitats.global
  52. 52.Group co-led by Fei-Fei Li suggests that AI safety laws should anticipate future risks | TechCrunchtechcrunch.com
  53. 53.Empowering Policymakers: Stanford HAI Trains Public Sector at Every Levelhai.stanford.edu
  54. 54.Stanford HAI's annual report highlights rapid adoption and growing accessibility of powerful AI systems - SiliconANGLEsiliconangle.com
  55. 55.Key findings from Stanford’s 2025 AI Index Report | IBMibm.com
  56. 56.Fei-Fei Li's World Labs comes out of stealth with $230M in funding | TechCrunchtechcrunch.com
  57. 57.The AI Dilemma: Growth Versus Existential Riskgsb.stanford.edu
  58. 58.Together raises $20M to build open source generative AI models | TechCrunchtechcrunch.com
  59. 59.‘The Godmother of AI’ says California’s well-intended AI bill will harm the U.S. ecosystem | Fortunefortune.com
  60. 60.Yoshua Bengio: California’s AI safety bill will protect consumers and innovation | Fortunefortune.com
  61. 61.The lawmaker behind California's vetoed AI bill, SB 1047, has harsh words for Silicon Valley | TechCrunchtechcrunch.com
  62. 62.McKinsey & Company Joins Corporate Affiliate Program at Stanford HAIhai.stanford.edu
  63. 63.Accenture Advances AI Research as Inaugural Member of Corporate Affiliate Program at Stanford Institute for Human-Centered Artificial Intelligencenewsroom.accenture.com
  64. 64.Stanford HAI Welcomes LVMH to the Corporate Affiliates Programhai.stanford.edu
  65. 65.AI pioneer Fei-Fei Li says AI policy must be based on ‘science, not science fiction’ | TechCrunchtechcrunch.com
  66. 66.Introducing The Foundation Model Transparency Indexhai.stanford.edu
  67. 67.The Foundation Model Transparency Index: What Changed in 6 Months?techpolicy.press
  68. 68.Stanford and The Fired Google AI Ethics Teamruth.substack.com
  69. 69.Stochastic Parrots 🦜 and Shaky Foundationsruth.substack.com
  70. 70.Timnit Gebru: Ethical AI Requires Institutional and Structural Changehai.stanford.edu
  71. 71.Stanford’s big AI bet: what if it busts?stanforddaily.com