Institute for Law & AI (LawAI)

Stated Theory of Change

LawAI's stated theory is that sound legal analysis will promote security, welfare, and the rule of law in the age of AI. The causal chain is:

1. AI governance requires specialized legal expertise that policymakers lack
2. LawAI produces rigorous legal research on AI governance challenges
3. This research directly informs legislation and regulation (through papers, consulting, and policy comments)
4. LawAI trains fellows who enter government and policy positions
5. The result is AI regulation that is more precise, constitutionally sound, and resistant to misspecification
6. Better regulation reduces catastrophic risk from AI while enabling beneficial innovation

The org's founder originally articulated this as "applied global priorities research through the means of law" -- using legal tools to address the world's biggest problems, with AI as the highest priority.

Revealed Theory of Change

The org's actions reveal a theory that is largely consistent with but more specific than the stated version. What they actually do:

Primary mechanism: Build intellectual infrastructure for AI regulation. The LFAI framework, governance misspecification concept, automated compliance model, and frontier model definitions work collectively provide a conceptual toolkit that policymakers can draw on. This is analogous to what the Brookings Institution or RAND does for defense policy -- you shape the conversation by being the people who have thought most carefully about the technical details.

Secondary mechanism: Serve as rapid-response legal analysts. The 25+ publications in 2025-2026, including commentary published within days of executive orders and state bills, positions LawAI as the go-to resource for real-time AI legal analysis. The Lawfare pipeline (through Rozenshtein) amplifies this enormously.

Tertiary mechanism: Train the next generation. The fellowship program is genuine field-building. There simply aren't enough lawyers who understand both AI capabilities and constitutional law. LawAI is trying to create that talent pool.

Where stated and revealed diverge: The org's research agenda includes forward-looking topics -- government-developed frontier AI, AI-enabled governance, AI rights as a strategic tool -- that go well beyond "help policymakers write better laws." These topics reveal the underlying x-risk concern that the neutral branding obscures. The rebrand from "Legal Priorities Project" (explicitly longtermist) to "Institute for Law & AI" (neutral think tank) is strategic -- it makes the work legible to policymakers who wouldn't engage with EA-branded research. This is arguably savvy rather than deceptive.

One notable absence in revealed priorities: The original 2021 vision included courts and litigation as a key pathway. Winter said: "going through the courts could be a viable option." This has been almost entirely deprioritized in favor of the legislative/executive approach. The $34K South Asia litigation grant from Open Phil appears to have been the last gasp of this thread.

Key Assumptions

Assumption 1: Legal frameworks matter for AI risk. If transformative AI arrives faster than legal systems can adapt, or if leading AI companies simply ignore legal constraints (as tech companies routinely do with privacy law), then legal scholarship has limited impact. Evidence for: Law has successfully constrained nuclear technology, genetic engineering, and other dangerous technologies. Evidence against: Law has failed to meaningfully constrain social media companies, cryptocurrency, or surveillance technology before harm occurred. Testable: Track whether any LawAI-influenced legislation actually constrains frontier AI development. If wrong: The entire theory of change collapses.

Assumption 2: Policymakers will listen to academic legal analysis. The consulting arm suggests LawAI believes direct engagement matters more than just publishing. But the evidence of specific policy influence is thin. Evidence for: The Founders Pledge assessment says LawAI "has provided valuable guidance to legislators and policymakers." California comments show direct engagement. Evidence against: No specific policy outcomes are publicly attributed to LawAI's work. The AI governance space is crowded with think tanks. Testable: Can LawAI point to specific legislative language or regulatory provisions that reflect their analysis? If wrong: The org produces high-quality scholarship that is cited in law reviews but ignored by legislatures.

Assumption 3: The legal-academic niche is genuinely neglected. Founders Pledge calls this "important and neglected." But Stanford Law, Harvard Law, and Georgetown (CSET) all have AI governance programs. Evidence for: Few orgs specialize specifically in the legal mechanics of AI regulation (as opposed to general AI policy). LawAI's work on proxy terms, compute thresholds, and model definitions is distinctively detailed legal analysis. Evidence against: GovAI covers similar territory. CSET at Georgetown does related work. The niche may be less neglected than claimed. If wrong: LawAI's marginal contribution is smaller than it appears.

Assumption 4: The LFAI framework is technically feasible. Law-Following AI requires AI systems to interpret and comply with open-textured natural-language laws. O'Keefe acknowledges the challenges: "we're not going to be able to have something completely theorized." Evidence for: Current LLMs already show impressive legal reasoning capabilities. AI companies are already instructing models to follow laws. Evidence against: Legal interpretation is contested even among human experts. Jurisdiction-dependent, evolving, and politically charged. O'Keefe himself says "very few people agree on like the underlying philosophy of law." Testable: Watch whether frontier AI companies implement anything resembling LFAI. If wrong: The flagship concept is interesting academic work that doesn't translate to practice.

Assumption 5: The $7.7M funder(s) are aligned with the mission. The org's independence depends on who provides 98% of its revenue. Evidence for: The org's history with SFF and Open Phil suggests EA-aligned funding. Evidence against: We literally don't know. If wrong: Funder misalignment could shape research priorities invisibly.

Strengths

Genuinely distinctive niche. No other org combines deep legal scholarship (law review publications), practical policy work (drafting model legislation), rapid commentary (Lawfare pipeline), and field-building (fellowship program) in the AI governance space. GovAI is more academic, CSET is more quantitative, and RAND is broader. LawAI fills a specific gap.

Intellectual ambition and honesty. The governance misspecification paper, AI rights paper, and automated compliance work show willingness to take novel positions and engage with uncomfortable ideas. The automated compliance podcast, where co-authors explicitly disagree and say so publicly, demonstrates intellectual culture that is rare in advocacy organizations.

O'Keefe as a unique asset. Having a former OpenAI policy researcher who left in the safety departures wave gives the org both insider knowledge and credibility. O'Keefe's ability to articulate AI safety concerns in legal language (AI henchmen, LFAI) bridges the gap between the technical alignment community and the legal/policy world.

Strategic positioning. The Lawfare connection provides massive distribution. The Cambridge affiliation provides academic credibility. The neutral branding makes the work legible to policymakers across the political spectrum. The fellowship program creates a long-term talent pipeline.

Financial stability. After nearly running out of money in 2023 ($262K in assets), the org now has $6M+ in reserves. This provides runway for multi-year research programs without the grant-chasing that distorts many think tanks.

Weaknesses and Risks

Funding opacity undermines the independence claim. An org that receives 98% of its revenue from unknown donor(s) and describes itself as "independent" is asking for trust it hasn't earned. Most comparable think tanks disclose at least category-level donor information. The absence of an annual report since 2022 compounds this.

No demonstrated policy impact. The org may be producing excellent legal scholarship that nobody reads. The California comments are the most visible engagement, but their effect is unknown. Without evidence of specific policy outcomes, the theory of change is aspirational rather than demonstrated.

Potential for Overton window distortion. By publishing in Lawfare and top law reviews, LawAI has outsized influence on how the AI governance conversation is framed. The LFAI concept in particular could shape policy debates in ways that benefit AI companies (by suggesting the problem is "make AI follow existing laws" rather than "restrict AI capabilities"). This may not be intentional, but the structural incentive exists.

Dependence on one funder. If the $7.7M came from a single donor (likely given the lack of fundraising infrastructure), the org is extremely vulnerable to that donor's continued support and potential shifting priorities.

Key person risk. The intellectual agenda is heavily shaped by O'Keefe. If he leaves, the org loses its most visible voice, its AI safety community credibility, and its connection to the OpenAI diaspora.

EU work appears underdeveloped. Despite having an EU Law team, published output is overwhelmingly US-focused. The EU AI Act is the most significant piece of AI regulation globally, and LawAI's analysis of it is not publicly visible.

Cross-References

Complementary: GovAI (Oxford-based, more academic and quantitative), CSET (Georgetown, more empirical and data-driven), IAPS (broader AI policy). LawAI provides the specifically legal analysis that these orgs lack depth in.

Related: CLTR (Centre for Long-Term Resilience, UK-based policy), Lawfare (national security law publication platform, not an org in the same sense). The Lawfare relationship is not competitive but symbiotic.

Ecosystem position: LawAI sits between the EA/AI safety community (where it originates) and the mainstream legal/policy world (where its work lands). It functions as a translation layer -- converting x-risk concerns into the language of constitutional law, statutory interpretation, and regulatory design. This positioning is inherently valuable but also creates the rebrand-as-opacity risk.

Staff network: O'Keefe from OpenAI, Koessler alumni to GovAI, Araujo alumni to Rethink Priorities, Salib connection to CAIS. The org is well-embedded in the AI safety governance network.

What Would Change This Assessment

Upward: (1) Evidence of specific policy wins -- a bill, regulation, or executive order demonstrably shaped by LawAI analysis. (2) Disclosure of the 2024 major donor(s), showing alignment with mission and absence of conflicts. (3) An external, critical evaluation from a credible source finding the work to be high-quality and impactful. (4) Fellowship alumni achieving senior policy positions and attributing their training to LawAI.

Downward: (1) Revelation that the major funder has industry ties (AI lab, tech company). (2) Evidence that research conclusions are shaped by funder preferences. (3) Key departures (especially O'Keefe) without adequate succession. (4) The LFAI framework being adopted by AI companies primarily as a liability shield rather than a genuine safety measure. (5) Continued failure to demonstrate any concrete policy impact after 2+ more years.

Self-Critique

Weakest claim: My assessment of the org's "genuine distinctiveness" may overweight the legal specialization. The AI governance policy space has many players, and it's unclear how much the specifically legal angle matters versus general policy analysis. I'm relying on Founders Pledge's "important and neglected niche" assessment without independent verification.

Potential bias: The evidence base is almost entirely self-generated by LawAI. Without external critics, I'm evaluating the org primarily through its own lens. The absence of criticism doesn't mean the work is above criticism -- it may mean the work is too niche to attract attention.

What a thoughtful disagreer would say: "LawAI is producing legal scholarship for legal scholars. The people who actually write AI regulation -- congressional staffers, agency rulemakers -- don't read law review articles. The consulting arm is the only part that matters for real-world impact, and we have zero evidence of its effectiveness. This is an academic project dressed up as policy influence."

Single weakest claim: That the LFAI framework will have meaningful real-world impact on how AI agents are governed. The concept is intellectually interesting but may be too novel and too dependent on contested legal philosophy to translate into enforceable regulation.

What would most change my view: Seeing the 2024 990 filing's Schedule I (if it exists) or the actual text of the major grant agreement. Understanding who provided the $7.7M and what conditions attached would fundamentally reshape the analysis. A critical academic response to the LFAI paper from a prominent legal scholar would also be very informative.