AI Policy Institute (AIPI)

Stated Theory of Change

AIPI's stated theory of change: demonstrate that the American public overwhelmingly supports AI regulation, use that data to arm policymakers with political cover, and build the political will to pass legislation. The underlying premise is that AI regulation is a democratic question, not primarily a technical one, and that public opinion is the underexploited lever.

Colson frames this explicitly: "Ultimately, policymakers are political actors, so the country needs an institution that can speak the language of public opinion sentiment. AIPI's mission is about being able to channel how Americans feel about artificial intelligence and pressure lawmakers to take action."

The c3/c4 structure operationalizes this: AIPI (c3) generates polling data, media coverage, and perceived expert status. AIPN (c4) takes that data to Congress through lobbying, congressional events, and direct engagement with legislative offices.

The implicit assumption: the bottleneck to AI regulation is not policy design (other orgs handle that) or technical understanding (policymakers will never be technical experts) -- it is political will. Politicians won't regulate AI until they believe voters want them to. AIPI's job is to prove they do.

Revealed Theory of Change

AIPI's actions largely match its stated theory, with some notable emphases and gaps:

Media-first strategy. AIPI hired a professional PR firm (Slingshot Strategies) from day one and pitched polls as exclusives to journalists. The org is more accurately understood as a media operation that uses polling as content than as a research institution that happens to get media coverage. This is not criticism -- it is a precise description of the strategy.

Bipartisan positioning is genuine. The Dean Ball collaboration, the IFS partnership, the AIPN team (Pentagon AI policy director, conservative and progressive Hill staffers), and the messaging that works across Bannon-to-Warren coalitions all demonstrate real bipartisan infrastructure. This is not cosmetic bipartisanship.

No policy substance. AIPI advocates for "regulation" and "guardrails" in general terms but has not proposed specific legislation, model bills, or regulatory frameworks. Compare to CAIP (model legislation the RAIA), IAPS (compute verification proposals), or CAIS (SB 1047 co-sponsorship). AIPI provides the "why regulate" but not the "what to regulate." This is a deliberate division of labor, but it means AIPI has no control over whether the policies that emerge from the political will it builds are actually good for AI safety.

The lobbying arm is where the action is. AIPN's $349K in lobbying spend, 24+ congressional events, and team of professional DC operatives represent real political infrastructure. The polling operation (AIPI) generates the ammunition; the lobbying operation (AIPN) fires it.

Zero engagement with technical AI safety. AIPI does not publish on alignment, interpretability, evaluations, or any technical topic. This is both a strength (staying in lane, maintaining credibility as a public opinion org) and a limitation (no ability to evaluate whether the policies AIPI supports would actually reduce AI risk).

Key Assumptions

1. Public opinion is the binding constraint on AI regulation.

• Evidence for: Politicians cite public opinion; AIPI's polls showing 72% support for slowing AI development create real political incentives. The NDAA preemption attempt was defeated partly due to public opposition.
• Evidence against: SB 1047 was vetoed despite AIPI's polls showing 70% support. Corporate lobbying, campaign donations, and executive connections often override public opinion. The American Prospect documented $4.2M in AI company donations to the California Democrats who opposed SB 1047.
• Testable: Track whether states/Congress pass AI regulation that can be causally linked to public opinion data.
• If wrong: AIPI's entire theory of change is ineffective -- it would mean politicians acknowledge public concern but are captured by industry interests regardless.

2. Polling data has sufficient shelf life to drive legislative change.

• Evidence for: AIPI's polls have been consistently cited for 2+ years. The numbers are remarkably stable across time, question framing, and survey house.
• Evidence against: Public attention is fleeting. COVID vaccine opinion shifted dramatically with changing circumstances. An AI boom that increases consumer benefit could erode anti-AI sentiment.
• Testable: Continue tracking AIPI's polling numbers over time.
• If wrong: AIPI would need to continually produce new polls to maintain relevance, which it appears to be doing.

3. Bipartisan support for AI regulation translates into legislative action.

• Evidence for: The anti-preemption coalition included Bannon, Warren, DeSantis, and MTG. 81% of battleground voters support guardrails.
• Evidence against: Bipartisan public support for gun control has not produced federal legislation for decades. The Trump administration actively opposes AI regulation. Congressional dysfunction prevents action on issues with even higher salience.
• Testable: Observe whether federal AI safety legislation passes in the next 2-3 years.
• If wrong: AIPI's polling becomes a database of irrelevant numbers -- proof that Americans want something they will never get.

4. "Regulation" and "guardrails" in general terms will produce policies that actually reduce AI risk.

• Evidence for: The specific bills AIPI supported (SB 1047, SB 53) had substantive safety provisions. AIPN's lobbying focuses on AGI preparedness, testing protocols, and nonproliferation.
• Evidence against: "Regulation" is vague enough to include anything from meaningful safety requirements to regulatory capture by incumbents. Without specifying what regulation looks like, AIPI could end up providing political cover for bad policy.
• If wrong: AIPI successfully builds political will for AI regulation, but the resulting legislation is co-opted by industry interests (a la Dodd-Frank) and fails to reduce risk.

Strengths

The polling data is real. Independent corroboration from IFS (conservative, 6,200 respondents), Pew, Economic Security Project, and the Center for AI Safety all show the same pattern: overwhelming public concern about AI and strong support for regulation. AIPI did not manufacture these numbers. It identified and amplified a genuine signal.

Bipartisan infrastructure is genuine and rare. Most AI governance orgs exist in a progressive/EA bubble. AIPI has built real connections across political lines -- Pentagon alumni, Republican Hill staffers, IFS partnership, Fox News commentary. In the current political environment, this is unusually valuable.

No industry funding conflicts. Unlike orgs dependent on OP (itself linked to tech wealth), AIPI appears free of economic ties to AI labs. The org can criticize Altman, Zuckerberg, and tech executives without biting the hand that feeds it. However, the "anonymous donors from tech and finance" caveat applies.

The collaborative poll with Dean Ball was genuinely smart. By partnering with an AI regulation skeptic on question design, AIPI demonstrated methodological confidence and created a result that even opponents have difficulty dismissing.

AIPN has real DC talent. The team's Hill experience is not cosmetic -- these are people who know how congressional offices work, how to brief staffers, and how to navigate the legislative process.

Weaknesses and Risks

Funding opacity undermines credibility. AIPI's insistence on anonymous donors in a space where funding transparency is the norm raises legitimate questions. "Dark money" criticisms from the Dossier are conspiratorial in tone but the underlying point -- that AIPI's funders are unknown -- is factually correct. This vulnerability could be exploited to discredit the org's polling.

No evidence of policy impact. SB 1047, AIPI's most prominent engagement, was vetoed. The NDAA preemption fight was won by a broad coalition, not AIPI alone. No legislator has publicly cited AIPI polling as the reason for their vote. The causal chain from polls to outcomes is undocumented.

Colson's Leverage Research background is a liability. Leverage Research's documented psychological experiments, cult-like characteristics, and traumatic effects on some members are well-established in public reporting. Colson's precise role is unclear, but his years at Leverage and his co-founding of Reserve (which recruited "basically a hundred percent" from Leverage initially) create a vulnerability to attack. If AIPI becomes prominent enough to attract serious opposition research, this will surface.

Single-person dependency. AIPI is functionally Daniel Colson's organization. No other person is named on the about page. If Colson leaves, AIPI ceases to exist as currently constituted. AIPN has broader team capacity, but even there, Colson is Executive Director.

The "American Affairs" critique is substantive. The argument that focusing on speculative x-risk distracts from demonstrable present harms (children's AI exposure, educational disruption, emotional manipulation) is gaining political traction. AIPI's polling asks about "catastrophic events" and "slowing AI" -- but the conservative policy movement is focusing on child safety and age-gating, which require different policy instruments than AIPI advocates for.

Polling without policy substance is hollow. Demonstrating that 80% of Americans want "guardrails" does not specify what those guardrails should be. AIPI creates demand for AI regulation but does not supply the regulatory content. This makes AIPI dependent on other orgs (IAPS, CAIS, Congress) to translate political will into policy, with no guarantee the resulting policies reduce risk.

Cross-References

Complementary to: IAPS (provides the policy substance AIPI lacks -- compute governance, legislative proposals, technical research), CAIS (provides the technical safety framing and legislative sponsorship), Encode Justice (grassroots youth AI advocacy with state-level action).

Different approach from: CAIP (direct congressional lobbying with model legislation; shut down for lack of funding), PauseAI (confrontational advocacy vs. AIPI's inside-game), MIRI (technical alignment research with no policy engagement).

Most similar to: ARI (Americans for Responsible Innovation) -- also a c3/c4 pair doing AI policy advocacy with bipartisan framing. ARI has OP funding and a broader issue set (including present harms); AIPI has no known institutional funders and focuses narrowly on polling.

Personnel overlap: Peter Wildeford's move from IAPS to AIPN creates direct intellectual cross-pollination between the OP-funded research orbit and AIPI's independent lobbying operation.

In the broader ecosystem: AIPI fills a specific niche -- the public opinion data provider for the AI safety policy movement. No other org does this at the same scale. If AIPI's polling disappeared, the movement would lose its primary evidence that the public supports what it advocates for.

What Would Change This Assessment

• If AIPI's funders were revealed to be AI lab-adjacent individuals: Would fundamentally change the incentive analysis and undermine the independence claim.
• If a legislator publicly cited AIPI polling as the decisive factor in supporting AI regulation: Would validate the theory of change with concrete evidence.
• If AIPI began publishing policy research or model legislation: Would expand the org's influence from political mandate to policy substance.
• If Colson's Leverage Research involvement were investigated further and found to include participation in the problematic debugging practices: Would create a serious credibility problem for the org.
• If public opinion on AI shifted toward optimism (e.g., due to visible AI benefits in healthcare/science): Would undermine AIPI's core asset -- consistent public concern.
• If SB 53 or successor legislation demonstrably improved AI safety: Would provide the first evidence that AIPI's polling-to-legislation pipeline produces safety outcomes.

Self-Critique

What sources should I have checked but didn't:

• Colson's Twitter/X activity, which likely contains real-time policy commentary and networking signals
• The TheStreet interview (403 error prevented access), which reportedly contained candid Colson quotes about "non-democratic tech leaders"
• AIPN's Manifund page in full (rate-limited), which may contain more detailed budget and impact information
• The AIPN RFI response PDF (could not extract text), which would reveal specific federal policy recommendations

Where this analysis may be biased:

• I may be over-weighting the Leverage Research connection. Colson was a young researcher there, not a leader, and the problematic practices may have been peripheral to his experience. Guilt by association is a weak analytical framework.
• I may be under-weighting the genuine value of the polling. In a world where politicians are captured by industry money, providing rigorous evidence of public opinion may be one of the few available counterweights.
• I may be applying an unfair comparison standard. AIPI is a small, young org being compared to IAPS ($15M+), CAIS (frontier lab connections), and CAIP (model legislation). Few orgs produce all of these outputs.

What a thoughtful disagreer would say: "AIPI identified and filled a genuine gap in the AI safety ecosystem. Before AIPI, the movement had technical researchers, policy analysts, and lobbying groups -- but nobody was systematically measuring and amplifying public opinion. AIPI's polls have become the standard data point cited by every AI regulation supporter. The SB 1047 veto doesn't prove the theory of change is wrong -- it proves that it takes more than one legislative cycle for public opinion to overcome industry lobbying. The Leverage background is irrelevant -- judge the org by its output, not its founder's resume from a decade ago. And the funding opacity, while unusual, may reflect Colson's personal wealth and small donor base rather than anything sinister."

My single weakest claim: That AIPI's polling has not produced policy impact. Impact in advocacy is genuinely difficult to measure, and the absence of documented impact is not the same as the absence of impact. AIPI's polls may have influenced countless staffers, journalists, and policymakers in ways that are never publicly attributed. The counterfactual -- what would the AI policy debate look like without AIPI's polling -- is unknowable.

What information would most change my view: Knowing AIPI's actual funders and budget. If AIPI is self-funded by Colson with a small budget, it's a lean, principled operation punching above its weight. If it's backed by specific tech billionaires with AI policy agendas, the incentive analysis changes entirely.