Center for Law & AI Risk (CLAIR)

Stated Theory of Change

CLAIR claims two interlocking paths from its activities to reduced AI risk:

Path A -- Intellectual: Develop legal frameworks that reshape how society governs AGI. The flagship idea is that granting AGIs private law rights (contract, property, tort) transforms the human-AI strategic interaction from a prisoner's dilemma into an iterated cooperative game. More broadly, CLAIR is producing scholarship across the entire spectrum of legal questions relevant to catastrophic AI risk -- tort liability, international cooperation, regulation design, corporate governance, tax incentives.

Path B -- Field-building: Assemble a community of legal academics who take catastrophic AI risk seriously and produce scholarship addressing it. The intended mechanism is analogous to how law-and-economics scholarship in the 1970s-80s reshaped antitrust policy: sustained intellectual effort creates an ecosystem of ideas that policymakers eventually draw upon.

Both paths depend on a critical assumption: that the transition to AGI will take long enough for legal scholarship and field-building to matter.

Revealed Theory of Change

CLAIR's actions are remarkably consistent with its stated theory of change. There is no significant gap between stated and revealed priorities:

• The co-directors are producing exactly the kind of scholarship they describe (high-quality law review articles applying game theory and economics to AI governance questions)
• The events (roundtable, writers' retreat, policy summit) are genuine field-building activities, not branding exercises
• Salib publishes regularly on Lawfare, translating academic ideas into accessible policy commentary
• The LunchGPT program extends to non-elite institutions, consistent with a genuine desire to grow the field broadly

One minor tension: the AI rights thesis gets the most attention and defines CLAIR's public identity, but the actual research network spans tort liability (Weil), systemic regulation (Arbel), governing AI agents (Kolt), and international cooperation (Salib/Goldstein). CLAIR is less "the AI rights org" than its most visible work suggests.

Key Assumptions

Assumption 1: Takeoff will be gradual enough for legal frameworks to matter.

• Evidence for: Current trajectory suggests incremental capability gains, not sudden FOOM. Agentic AI is arriving in stages. Regulatory infrastructure takes years to build.
• Evidence against: Rapid capability gains in 2023-2025 exceeded most predictions. If AGI arrives in 2-3 years, academic scholarship started in 2024 will not have time to influence policy.
• Testable: Yes -- directly falsified by sudden capability jumps or AI systems that bypass legal authority.
• If wrong: CLAIR's work has primarily retrospective/historical value, not safety value.

Assumption 2: AIs will be agentic enough to participate in game-theoretic interactions but not so powerful that cooperation is unnecessary.

• Evidence for: Current frontier models show increasing agency. The "moderate capability" window is the most likely near-term scenario.
• Evidence against: The window could be very narrow, or the transition through it could be so fast that legal frameworks cannot adapt.
• Testable: Partially -- depends on empirical measurements of AI capability and the speed of capability growth relative to institutional adaptation.
• If wrong: The AI rights thesis specifically doesn't apply. Other CLAIR research (tort liability, regulation) may still be useful.

Assumption 3: Legal academia can influence AI policy in time.

• Evidence for: Historical precedent (law-and-economics reshaping antitrust, Khan's student note reshaping FTC). Virginia Law Review publication signals mainstream legal acceptance. Fordham and Oxford engagement.
• Evidence against: The antitrust analogy took decades. If timelines are short, there is no time for the slow diffusion of ideas through law reviews. Direct policy advocacy (like what the Institute for Law & AI does) may be needed instead.
• Testable: Track whether CLAIR's ideas appear in legislation, regulatory proposals, or judicial opinions.
• If wrong: The field-building theory of change fails, though the intellectual contributions may still be valuable through other channels.

Assumption 4: Comparative advantage will sustain human-AI trade.

• Evidence for: Economic theory supports the existence of gains from trade under very general conditions. Salib's Alice-the-tax-lawyer example illustrates comparative advantage even with absolute disadvantage.
• Evidence against: Transaction costs may dominate, wages could be sub-subsistence, communication barriers may prevent efficient trade. Barnett/Epoch research suggests wages could fall below subsistence.
• Testable: Yes -- this is an empirical question about the economics of human-AI interaction.
• If wrong: The iterated positive-sum game that makes AI rights valuable generates insufficient surplus to prevent defection.

Strengths

1. Intellectual novelty in an underexplored space. CLAIR is one of very few organizations applying serious legal scholarship to catastrophic AI risk. The AI safety field is heavily dominated by technical alignment and control research. CLAIR addresses a genuine blind spot.

2. Unusual candor about limitations. Salib openly discusses failure modes of his own proposal. Weil published "The Limits of Liability" identifying five categories of problems tort law cannot solve. Intellectual honesty is rare and valuable.

3. Lean, independent operating model. University-hosted, no apparent dependence on large funders. This gives CLAIR unusual freedom from the incentive distortions that affect many AI safety orgs (e.g., not beholden to the labs they might need to criticize).

4. High-quality publication venue. The flagship paper was accepted by Virginia Law Review -- a top-10 US law journal. This is a credibility signal for the legal academy.

5. Bridge-building between communities. CLAIR brings AI safety thinking to legal academia, which has enormous downstream influence on regulation. The roundtable attracted ~25 legal scholars who might otherwise never engage with x-risk.

6. Salib's pragmatism. The "Bird in the Hand" piece shows Salib is not just an ivory-tower theorist -- he is willing to support imperfect regulation now rather than waiting for perfect regulation that may never come. This practical instinct is valuable.

Weaknesses and Risks

1. Timeline vulnerability. If AGI arrives soon and takeoff is fast, CLAIR's academic field-building approach will not have produced usable policy frameworks in time. This is the existential risk to CLAIR's theory of change.

2. Extremely small scale. Two co-directors, no staff. The entire output depends on Salib's productivity and the willingness of a handful of affiliated scholars to contribute. One person's career disruption could stall the project.

3. No verified policy pathway. The antitrust analogy is encouraging but took decades. CLAIR has no concrete mechanism for translating scholarship into policy faster than historical precedent suggests. The Policy Summit could be an attempt, but details are absent.

4. Funding opacity. We cannot assess sustainability, independence, or growth trajectory because we know nothing about CLAIR's funding. This is the single biggest informational gap.

5. Limited engagement with technical AI safety community. CLAIR operates almost entirely within legal academia and Lawfare. It has no LessWrong/AF presence, no collaboration with technical alignment researchers (beyond CAIS affiliation), and no engagement with the policy advocacy organizations that might carry its ideas into practice.

6. The AI rights thesis is ahead of its time. Even Salib acknowledges it sounds radical. If AGI arrives and crisis forces action, policymakers may reach for simpler tools (bans, mandatory control, nationalization) rather than the nuanced "small-l liberalism for AIs" that CLAIR proposes.

Cross-References

• Center for AI Safety (CAIS): Organizational parent. CAIS funded and hosted the founding workshop. Salib is CAIS advisor. The relationship appears to be supportive rather than directive.
• Institute for Law & AI (LawAI): Sister organization. Harvard-affiliated think tank with overlapping personnel (Salib, Frazier). More focused on direct policy advising than CLAIR's academic field-building.
• Forethought: MacAskill-founded longtermist research org. Salib is Visiting Senior Scholar.
• AI Frontiers: Media platform where Salib and Weil publish accessible pieces.
• Complementary to technical alignment orgs (ARC, MIRI, Redwood, Anthropic's alignment team): CLAIR's work is downstream -- it assumes the "moderate alignment" scenario and asks how to manage it institutionally rather than technically.
• Distinct from policy advocacy orgs (Center for AI Policy, GovAI): CLAIR is upstream -- producing scholarship that could feed into policy, not doing policy advocacy directly.

What Would Change This Assessment

Strong positive update:

• CLAIR research cited in legislation, regulatory proposals, or judicial opinions
• Salib testifies before Congress or advises an executive agency
• Computational game theorists formalize the AI rights model and find wide parameter regions where it works
• A major AI company (e.g., Anthropic, OpenAI) incorporates CLAIR's framework into their governance thinking

Strong negative update:

• Fast takeoff renders legal frameworks irrelevant before they can be implemented
• CLAIR loses both co-directors or key collaborators
• The AI rights idea is formally refuted by game theorists
• CLAIR is revealed to be primarily funded by an entity with conflicts of interest (e.g., an AI company)

Self-Critique

What sources should I have checked but didn't?

• Arbel's personal site (battleoftheforms.com) yielded only 30 words; his views are underrepresented
• CLAIR's blog page failed to fetch; may have relevant content
• The actual SSRN papers (full text, not just summaries) -- the Lawfare pieces and podcast are more accessible but less rigorous

Where is this analysis potentially biased?

• I may be overly sympathetic to the intellectual project because the AXRP interview showed Salib being unusually candid and thoughtful. Candor is not the same as being correct.
• I may underweight the "timeline vulnerability" objection because I spent more time reading CLAIR's arguments for its relevance than reading arguments that timelines are short.

What would a thoughtful person who disagrees say?

• "This is academic theorizing about a scenario (cooperative AGIs) that presupposes we fail at alignment but succeed at building AGIs that are rational enough to play game theory. That is an extremely narrow slice of outcome space, and building legal infrastructure for it is a poor allocation of scarce resources."
• "Salib's comparative advantage argument is hand-waving. The Epoch/Barnett work on AI economics suggests human wages go to near-zero, which means the cooperation surplus that drives the whole model is negligible."

What is my single weakest claim?

• My assessment that CLAIR's lean operating model is a strength rather than a weakness. It could instead mean the org will never grow beyond a hobby project for two professors and will have no impact at scale.

What information would most change my view?

• Evidence that CLAIR's research has influenced a specific policy debate (positive update) or detailed analysis showing the parameter space for beneficial AI rights is negligibly small (negative update).