Center for AI Policy (CAIP)

Stated Theory of Change

CAIP's theory of change had three links:

1. Build congressional relationships and credibility through persistent engagement: hundreds of meetings, briefings, endorsements, and media presence.
2. Develop ready-made legislative frameworks (the RAIA model legislation) that can serve as a template when Congress is motivated to act.
3. Wait for a triggering event -- either a grassroots movement reaching critical mass, or a "warning shot" disaster creating a legislative window -- and then deploy the pre-positioned relationships and policy proposals to shape the response.

The explicit reference points were the PATRIOT Act (post-9/11) and Dodd-Frank (post-2008 financial crisis): large, complex legislation that passed quickly because the crisis created political will, and pre-existing policy frameworks existed to channel that will.

The mechanism for ongoing safety improvement was mandatory private audits backed by a federal agency with power to block deployment, combined with liability reform, hardware monitoring, and emergency powers. CAIP argued these tools would "pick up a lot of the low-hanging fruit around reducing the harms from near-term, accidental, and/or half-baked deployments."

Revealed Theory of Change

CAIP's actions were largely consistent with their stated theory. The 406 congressional meetings, 8 briefings, 12+ bill endorsements, and 3 accepted bill edits all support the "build relationships and credibility" narrative. The model legislation was a genuine and serious policy proposal (confirmed by Zvi's analysis). The grassroots organizing (PAN, Demo Day) was a real attempt to build constituency.

Where revealed and stated theories diverge:

1. The urgency framing was unrealistic. The founding post (August 2023) talked about "harnessing the current energy" and passing legislation during "this policy window." The 118th Congress passed no AI safety legislation. The "policy window" closed without CAIP achieving its primary objective. By the time of the shutdown, the theory had shifted to "wait for a warning shot" -- a much more patient, uncertain bet.

2. The dual-track strategy may have diluted impact. CAIP simultaneously pushed ambitious model legislation AND endorsed modest incremental bills. This is strategically sensible but means their limited resources were spread across objectives with very different timelines and political dynamics.

3. Internal priorities shifted over time. The org was originally founded by a technical AI safety researcher (Larsen) with x-risk convictions, but evolved to be led by a lawyer (Green-Lowe) with a DC advocacy strategy. The co-founders who departed may represent the tension between the x-risk motivation and the pragmatic DC approach. The "revealed" theory of change was more conventional lobbying than the founding vision suggested.

Key Assumptions

Assumption 1: Congress can act fast enough to matter.

• Evidence for: Historical examples (PATRIOT Act, Dodd-Frank). SB 1047 passed California Senate quickly. Congress does sometimes move fast under pressure.
• Evidence against: The 118th Congress was "historically slow and unproductive." AI development timescales may be shorter than legislative timescales. CAIP's own co-founder updated that congressional advocacy was intractable on relevant timelines.
• Testable: Yes -- we would need to see Congress pass mandatory AI safety requirements within the next 2-3 years.
• If wrong: CAIP's entire theory of change collapses. All the relationship-building and model legislation would be irrelevant if AI reaches transformative capability before Congress acts.

Assumption 2: Mandatory audits would meaningfully reduce catastrophic risk.

• Evidence for: Current voluntary audits are rushed and incomplete (Green-Lowe). Mandatory audits improve auditor bargaining power. Similar regimes exist in nuclear, aviation, and finance.
• Evidence against: Zvi's critique: the audit criteria are too vague and subjective. Auditing for existential risk is qualitatively different from auditing for product safety. Alignment failures may not be detectable by any currently known evaluation. The RAAIA's own "extremely high concern" tier is essentially a TBD.
• Testable: Partially -- you could evaluate the detection rate of red-teaming exercises.
• If wrong: Mandatory audits become security theater, creating false confidence while existential risks continue to accumulate.

Assumption 3: Pre-positioned policy proposals will be adopted during a crisis.

• Evidence for: Historical precedent (existing frameworks were adopted in prior crises). CAIP built the relationships and credibility needed.
• Evidence against: CAIP is now shut down. The relationships and institutional knowledge are dissipating. Other orgs (ARI, CAIS AF) may fill the vacuum but with different proposals. A crisis may produce ad hoc legislation rather than adopting a pre-existing framework.
• If wrong: CAIP's main legacy -- the model legislation -- becomes an artifact that influences future thinking but is never directly adopted.

Assumption 4: The funding ecosystem will eventually correct its bias against advocacy.

• Evidence for: The shutdown post is a strong call to action. SFF made "multiple bets" on advocacy. OP funded ARI.
• Evidence against: The structural barriers (c4 non-deductibility, DAF restrictions) are not going away. The same funders who rejected CAIP in 2025 are still the primary AIS funders.
• If wrong: The advocacy gap persists, and the AIS movement continues to produce research that never reaches legislative form.

Strengths

1. CAIP was genuinely unique. No other organization combined a 501(c)(4) status, Congressional focus, mandatory safety legislation, and grassroots network. ARI comes closest but has different political strategy and OP funding ties.

2. The model legislation was serious. Zvi's 10,000-word review confirmed this was not a vanity project. The bill addressed real policy questions with real specificity, even if imperfect. No other AIS org has produced comparable legislative text.

3. Bipartisan positioning was smart. In an era of increasing partisanship on tech issues, CAIP deliberately cultivated both parties and adapted their proposals accordingly (2025 RAIA was "more Republican-friendly"). This is the right approach to avoid the climate failure mode.

4. No AI company money. This gave CAIP moral authority and independence that industry-funded orgs lack. Their willingness to call Altman's proposals "vaporware" demonstrated real independence.

5. High output per dollar. ~$1.6M/year produced 406 meetings, model legislation, 8 briefings, 20 events, and demonstrable congressional engagement. This is remarkably efficient for DC advocacy.

Weaknesses and Risks

1. The theory of change depends on events outside CAIP's control. If no "warning shot" comes before transformative AI, or if the warning shot is too large for normal government to function, the entire investment is wasted. This is the fundamental weakness.

2. CAIP's own co-founder concluded the approach was intractable. Thomas Larsen's departure is the single most damaging data point. If the person who co-founded the org specifically to do congressional advocacy decided it would not work, external observers should update heavily.

3. The c4 structure was a trap. It enabled unrestricted lobbying but cut off the funding channels that sustain AI safety organizations. CAIP never established a sister c3, which was a critical strategic error. ARI has both.

4. Impact is genuinely unmeasurable. CAIP themselves said: "we don't really know" whether the work was impactful. Activity metrics (meetings, endorsements) do not translate to outcome metrics (legislation passed, risk reduced). The 118th Congress passed no AI safety legislation.

5. Institutional knowledge is dissipating. With shutdown, the relationships, political intelligence, and legislative drafting expertise are dispersing to other organizations or leaving the field entirely. The sunk cost of relationship-building cannot be recovered.

6. Resource asymmetry is crushing. CAIP spent ~$500K on lobbying over two years. Meta spent $24M in one year. The structural imbalance means advocacy orgs need to be orders of magnitude more efficient than their opponents, which may be impossible.

Cross-References

Americans for Responsible Innovation (ARI): CAIP's closest comparator. ARI is the best-funded advocacy org in the space (large OP grants for both c3 and c4 work), with a broader focus (including present harms, not just x-risk). ARI represents the "well-funded, mainstream" approach to AI advocacy vs. CAIP's "scrappy, x-risk-focused" approach. ARI's survival while CAIP died suggests that OP-fundable strategy wins.

Survival and Flourishing Fund (SFF): CAIP's primary funder and the only major grantmaker willing to bet on c4 advocacy. SFF's retreat from c4 funding (noted in the funding appeal) was a death blow.

CAIS Action Fund: CAIS AF focuses on the national security angle with an emphasis on chip security. Funded by SFF. Occupies an adjacent but distinct niche.

PauseAI: Occupies the "more radical" end of the advocacy spectrum that CAIP deliberately avoided. PauseAI represents the approach CAIP rejected: pushing for a pause/moratorium rather than mandatory audits.

IAPS / GovAI / CSET: Research-focused governance orgs that produce the kind of analysis CAIP argued was necessary but insufficient. CAIP's shutdown highlights the tension between research (fundable, low-risk) and advocacy (underfunded, higher-risk).

Encode: Grassroots-focused advocacy with state-level work. Co-sponsored SB 1047. Funded by FLI and SFF.

What Would Change This Assessment

• If a CAIP 2.0 launches and succeeds in raising sustained funding: This would suggest CAIP's failure was execution-specific rather than structural.
• If Congress passes mandatory AI safety legislation within 3 years: This would validate the core theory of change, even if CAIP itself did not survive to see it.
• If the RAIA framework is adopted as a template by a future legislative effort: This would suggest CAIP's model legislation was the most durable contribution, even post-shutdown.
• If the advocacy funding gap closes significantly: This would suggest CAIP's failure was a symptom of a broader dysfunction that is being corrected.
• If Thomas Larsen publicly revises his view on congressional tractability: This would undercut the strongest negative signal.

Self-Critique

What is weakest in this analysis:

• I may be over-weighting Thomas Larsen's departure as a negative signal. People leave organizations for many reasons (personal fit, disagreement about strategy, career opportunities). His departure does not necessarily mean the theory of change is wrong -- it means one person concluded it was not the right fit.
• I have no independent verification of CAIP's claimed outputs (406 meetings, 3 bills edited). These come entirely from CAIP's own statements.

What a thoughtful disagreer would say: "CAIP failed because of execution problems (hiring, fundraising, management) and structural barriers (c4 status), not because the theory of change was wrong. The fact that Congress has not yet passed AI safety legislation does not mean it never will. CAIP was a first attempt at a very hard problem, and first attempts often fail. The correct response is to learn from their mistakes and try again -- not to abandon advocacy."

What I missed:

• I was unable to read Jason Green-Lowe's full 7-part EA Forum sequence, which may contain the most detailed and sophisticated articulation of the case for AI safety advocacy. Posts 2-7 are on EA Forum (inaccessible).
• I have no information about how congressional staffers themselves evaluated CAIP. All feedback is from the AI safety community, not from the policy professionals CAIP was trying to influence.
• I do not have CAIP's 2024 Annual Report (PDF), which may contain more detailed impact claims.

Single weakest claim: My assertion that "the theory of change depends on events outside CAIP's control" is both true and unfair -- virtually all policy advocacy depends on political windows that advocates cannot control. The question is not whether CAIP controlled the timing, but whether the pre-positioning was valuable. I cannot definitively answer that question.