Stanford HAI

Stated Theory of Change

HAI's stated theory: AI risk is best managed through interdisciplinary academic research, evidence-based governance, transparency measurement, and compute democratization. The mechanism is: produce authoritative public goods (AI Index, FMTI, HELM) that inform policymakers and the public --> train legislators and staffers directly --> build infrastructure (NAIRR) that prevents total industry dominance --> the resulting ecosystem produces safer AI development.

Li articulates this as a "human-centered" philosophy: "AI is a tool, and its values are human values." This implies the problem is getting human institutions right (policy, oversight, education, diversity) rather than solving technical alignment problems in the AI itself. The theory is essentially: good governance + informed publics + accessible technology = safe AI outcomes.

Revealed Theory of Change

HAI's actions reveal a more complex picture:

What they actually prioritize: Measurement tools (AI Index, FMTI, HELM), policy engagement (NAIRR, congressional training), and interdisciplinary research funding. These are real, substantial contributions. The AI Index is genuinely the most comprehensive survey of AI progress. NAIRR is a concrete policy achievement that addresses a real structural problem. HELM Safety does useful safety evaluation work.

Where actions diverge from stated theory:

1. Li's SB 1047 opposition aligned with every major corporate partner, then she shifted 7 months later to co-lead a working group recommending similar measures. This suggests positions are influenced by institutional context, not just principled analysis.
2. Li's personal financial interests ($5B World Labs) create a structural incentive to oppose regulation that might slow AI development, even as HAI studies AI's societal impact.
3. FMTI's own results show transparency declining, which means HAI's measurement-and-shame theory of change may not actually work -- yet HAI hasn't pivoted to stronger mechanisms (regulation, binding commitments).
4. The Project Maven emails reveal that Li's private concern was brand damage, not ethics. Her public statement about principles was constructed after the fact.
5. HAI distributes $50M+ in research funding broadly across Stanford, with no visible prioritization of catastrophic risk research. The safety-relevant work (Liang, Barrett, Hashimoto) is funded by Open Philanthropy, not by HAI's corporate donors.

Revealed theory of change: HAI functions as a legitimacy bridge between Stanford, industry, and government. It provides academic credibility to the idea that AI can be governed through voluntary transparency and informed policy, without binding regulation or fundamental changes to the industry structure. Whether intentional or not, this serves industry interests by channeling governance impulses into measurement rather than restriction.

Key Assumptions

1. Measurement changes behavior.

• Evidence for: Companies did initially improve on FMTI (58 average in 2024). Some companies cite FMTI as motivation for transparency improvements.
• Evidence against: FMTI 2025 shows transparency declined to 40 average. Meta and OpenAI, initially first and second, are now last and second-to-last. The measurement is not preventing backsliding.
• Testable: Yes -- track FMTI scores over time. If the next 2 editions show continued decline, the measurement theory is falsified.
• If wrong: HAI's primary theory of change (measurement + public pressure = better behavior) fails, requiring stronger mechanisms (regulation, binding commitments).

2. Interdisciplinary research produces better AI governance.

• Evidence for: NAIRR was developed through interdisciplinary collaboration. The AI Index incorporates economic, social, and policy perspectives.
• Evidence against: No clear example where HAI's interdisciplinary work has changed a company's actual AI development practice (as opposed to its documentation).
• Testable: Harder to test. Would require tracing specific policy outcomes to HAI research inputs.
• If wrong: The "interdisciplinary" framing adds academic legitimacy but doesn't change what actually gets built.

3. "Human-centered" framing is compatible with catastrophic risk reduction.

• Evidence for: Li's March 2025 working group report explicitly acknowledged that "extreme risks" deserve policy attention. Some HAI-affiliated researchers (Liang, Barrett, Hashimoto) do alignment-adjacent work.
• Evidence against: Li's consistent deprioritization of x-risk in favor of near-term harms. HAI's SB 1047 opposition actively worked against safety regulation. The "science not science fiction" framing delegitimizes x-risk concerns without engaging them.
• Testable: Partially. If HAI's framing leads policymakers to focus exclusively on bias/fairness while ignoring catastrophic scenarios, that would be evidence against.
• If wrong: HAI is not just neutral on catastrophic risk but actively harmful, by occupying the "responsible AI" space while channeling attention away from the most important risks.

4. Academic independence is maintained despite corporate funding.

• Evidence for: HAI publishes FMTI scores that embarrass corporate partners. Schaake's "Tech Coup" critiques industry power from within HAI.
• Evidence against: Project Maven emails, SB 1047 alignment with industry, advisory council composition, Stanford's institutional financial dependence on AI.
• Testable: Yes -- look for cases where HAI takes positions directly contrary to its corporate funders' interests on high-stakes issues.
• If wrong: HAI functions as a sophisticated form of industry influence-washing, providing academic cover for industry-preferred governance approaches.

Strengths

1. Genuine public goods production. The AI Index and FMTI, whatever their limitations, are the most comprehensive publicly available measures of AI progress and transparency. They create information that didn't exist before. The AI Index is cited in virtually every serious policy discussion.

2. NAIRR is a real achievement. Taking a concept from a 2019 essay to federal legislation and an NSF pilot in 5 years is substantial policy work. It addresses a genuine structural problem (industry-academia compute gap) that most safety-focused orgs don't touch.

3. Stanford's institutional power. HAI can convene policymakers, train congressional staffers, and place students in DC offices at a scale no standalone safety org can match. When HAI says something, members of Congress listen in a way they wouldn't for MIRI or CHAI.

4. CRFM/Liang's technical work. HELM Safety, FMTI, and Liang's alignment research are genuinely useful. The dual funding structure (OP funds the safety work, industry funds the institution) inadvertently creates some of the most practically useful safety evaluation infrastructure in the field.

5. Willingness to publish uncomfortable findings. FMTI 2025 showing declining transparency is embarrassing for the industry-funded institute. Publishing it anyway suggests some genuine independence.

Weaknesses and Risks

1. The "human-centered" framing may actively harm catastrophic risk reduction by redefining "AI safety" to mean bias, fairness, and workforce disruption while delegitimizing x-risk concerns as "science fiction." Policymakers who learn about AI from HAI may conclude that the serious risks are near-term and the catastrophic risks are speculative. This is exactly what happened with SB 1047.

2. Industry capture is structural, not intentional. HAI's leadership may genuinely believe their positions are correct. But when your advisory council includes the Alphabet board chair, your compute comes from Google and Microsoft, your founding members pay $5M/year, and your co-director runs a $5B AI startup, the incentive alignment with industry is overwhelming. It doesn't require anyone to act in bad faith.

3. The theory of change may be circular. HAI measures transparency, finds it declining, publishes the finding, but has no mechanism to compel change. If measurement alone doesn't change behavior (and the data suggests it doesn't), then HAI's primary output is documentation of failure rather than prevention of harm.

4. Li's dual role is an unresolved conflict. Running a $5B AI capabilities company while co-directing an institute meant to guide AI responsibly is a conflict that no values statement can resolve. World Labs builds the technology; HAI is supposed to ensure it benefits society. The same person controls both.

5. Absence from the safety community. Zero engagement with the AI safety community's institutions, media, or intellectual infrastructure. No presence on LW, EA Forum, 80K Hours, or any x-risk-oriented platform. HAI and the alignment community essentially operate in parallel universes, which means HAI's substantial policy influence is not informed by the most rigorous thinking about catastrophic risk.

Cross-References

• vs. CHAI (Berkeley): CHAI takes a technical approach to alignment under Stuart Russell's "provably beneficial AI" framework. HAI takes an institutional approach under Li's "human-centered AI" framework. These are fundamentally different theories of change -- CHAI says the AI itself must be aligned; HAI says human institutions must govern AI wisely. Both could be right, but HAI's approach doesn't address the case where AI systems become too capable for human institutions to govern.

• vs. CSET (Georgetown): Both do AI policy work, but CSET focuses on national security and great power competition while HAI focuses on governance, transparency, and compute access. HAI's policy influence is broader but shallower.

• vs. Coefficient Giving/Open Phil: OP funds specific safety-relevant researchers at Stanford (Liang, Hashimoto, Potts, Barrett) rather than HAI as an institution. This suggests OP sees the researchers as valuable but HAI's institutional direction as not aligned with safety priorities.

• vs. FLI: Both produce reports and engage with policy. FLI's AI Safety Index focuses specifically on frontier lab safety practices. HAI's FMTI focuses on transparency more broadly. FLI is explicitly concerned about catastrophic risk; HAI is explicitly not.

• Complementary role: HAI's policy infrastructure (congressional bootcamps, NAIRR, AI Index) serves a function no x-risk-focused org fills. If the safety community could influence what HAI teaches policymakers, HAI's institutional power could amplify safety concerns. Currently, it doesn't.

What Would Change This Assessment

1. HAI explicitly adds catastrophic/existential risk to its research agenda with dedicated funding, not just individual researchers working on OP grants. This would signal that "human-centered" and "safety-focused" are not in tension.

2. HAI supports binding AI regulation -- not just transparency requirements but liability provisions, compute governance, or deployment restrictions. This would demonstrate independence from industry.

3. FMTI scores start improving again, demonstrating that measurement can change behavior without regulation. This would validate HAI's theory of change.

4. Li publicly engages with x-risk arguments on their merits rather than dismissing them as "science fiction." Specifically, if she responded to Russell, Bengio, or Hinton's technical arguments about why advanced AI might be catastrophically dangerous.

5. HAI publishes its own budget and funding sources. This would resolve the transparency hypocrisy and allow evaluation of industry dependence.

6. Li steps back from World Labs or establishes a formal firewall between her commercial interests and HAI's governance work.

Self-Critique

What sources should I have checked but didn't?

• HAI's annual report PDFs (2023, 2024), which were not extractable. These likely contain the most detailed operational data.
• The Nature article on "Focal points and blind spots of human-centered AI" -- a potentially valuable academic critique that failed to download.
• Direct interviews or publications from HAI's corporate founding members about what they get from the partnership.

Where is this analysis potentially biased?

• I may overweight the Project Maven emails because they're dramatic. Li's behavior in 2017 may not reflect her current views.
• I may overweight the absence of x-risk focus because my analytical frame (AI safety) makes that absence salient. HAI's near-term work on bias, workforce, and healthcare has real value that I may underweight.
• The "industry capture" framing is itself a frame. It's possible HAI's positions are genuinely independently derived and happen to overlap with industry because the same evidence leads to the same conclusions.

What would a thoughtful person who disagrees say? "HAI is doing exactly what a responsible academic institute should do: producing evidence, measuring transparency, training policymakers, and addressing the risks we can actually observe right now. The x-risk community's fixation on speculative scenarios is itself a form of distraction from real, documented harms. Li is right that 'science, not science fiction' should guide policy. And the fact that HAI's positions sometimes align with industry doesn't prove capture -- it might just mean that some industry positions are correct."

What's my single weakest claim? That HAI's "human-centered" framing actively harms catastrophic risk reduction. It's possible the framing is simply neutral -- it addresses a different set of problems. The harm thesis requires showing that policymakers who learn from HAI are less likely to support x-risk regulation than they would be without HAI's influence, which I cannot demonstrate with available evidence.

What information would most change my view? HAI's internal budget breakdown showing what percentage comes from corporate vs. philanthropic vs. federal sources. If corporate funding is actually a small fraction and the Hoffman-Yee and federal grants dominate, the industry capture thesis would need substantial revision.