Americans for Responsible Innovation (ARI)

Stated Theory of Change

ARI claims that building "regulatory muscle" -- government capacity to regulate AI effectively -- is the best response to deep uncertainty about AI's trajectory. Rather than betting on a specific risk timeline, they argue policymakers should build institutions, standards, and legal frameworks that work across scenarios. The mechanism is bipartisan lobbying plus grassroots mobilization to pass targeted legislation (AISI authorization, whistleblower protections, export controls, incident reporting, liability frameworks) and block industry-favorable preemption.

The distinctive intellectual contribution is Carson's claim that near-term and long-term AI risks are causally linked through institutional degradation: addressing deepfakes, child safety, and algorithmic bias builds the institutional capacity needed to later address misalignment. This rejects the common framing where near-term harms and x-risk are competing priorities.

Revealed Theory of Change

ARI's actions are broadly consistent with its stated theory, with some informative emphases:

Lobbying is the core product. Team composition (~27 staff, heavily drawn from Congressional offices), the highest salary going to the SVP of Government Affairs ($438K vs. ED's $253K), the multi-entity legal structure (c4 + c3 + PAC + Carey Committee), and the connection to Public First Action super PACs all point to an organization built for political influence. This is not a research org that also lobbies -- it is a lobbying operation with a research support function.

Near-term wins build political capital. ARI's two biggest wins (TAKE IT DOWN Act, moratorium defeat) are near-term harm issues, not emerging risk. This is strategically smart: you cannot pass frontier AI governance legislation without first proving you can pass any legislation. But it means ARI's actual contribution to long-term AI risk reduction is currently indirect -- building the political infrastructure and relationships that could later be deployed on harder issues.

The national security angle is the on-ramp. Export controls, chip policy, and GAIN AI Act work positions ARI where bipartisan consensus is strongest. This is the path of least political resistance and maximum near-term credibility.

Carson's parallel political operation is the big strategic bet. Public First Action ($20M+ from Anthropic, targeting $50-75M) aims to elect pro-regulation candidates in 2026 midterms. If successful, this changes the Congressional composition ARI lobbies -- a second-order theory of change that could be far more impactful than any individual bill.

Key Assumptions

1. Regulatory muscle built for near-term harms transfers to frontier AI governance.

• Evidence for: Historical precedent (FDA, FAA, NHTSA were built incrementally). NIST/AISI is genuinely dual-use infrastructure.
• Evidence against: Near-term AI regulation (deepfakes, child safety) may be structurally different from frontier governance (training run oversight, alignment verification). The regulatory muscle for one may not fit the other.
• Testable: Within 2-3 years. If AISI is authorized and funded, does it develop frontier governance capacity?
• If wrong: ARI has built a successful near-term harm advocacy org but not contributed to reducing catastrophic AI risk.

2. Bipartisan centrism is the right strategy for AI safety.

• Evidence for: TAKE IT DOWN passed with Cruz-Klobuchar sponsorship. Moratorium defeated 99-1. ARI PAC gives 58% R / 42% D. The evidence suggests bipartisan approaches actually pass.
• Evidence against: The hardest AI governance questions (compute monitoring, training run restrictions, mandatory safety cases) may not have bipartisan solutions. The Overton window may need moving, not navigating.
• If wrong: ARI passes popular, moderate legislation but never reaches the frontier governance measures that matter most for catastrophic risk.

3. The spending asymmetry can be overcome through political strategy.

• Evidence for: ARI and allies defeated the moratorium despite being outspent 5:1+. Polling shows 80% of Americans support AI safety rules. Public opinion is on ARI's side.
• Evidence against: $125M+ from Leading the Future vs. $50-75M target for Public First Action. Big Tech has persistent structural advantages (relationships, revolving door, expertise). The moratorium was a particularly egregious proposal -- more subtle industry-favorable provisions may be harder to stop.
• If wrong: ARI achieves symbolic wins while industry writes the substantive rules.

4. Carson's dual role (ARI president + Public First Action co-chair) is manageable without conflict of interest.

• Evidence for: ARI's no-corporate-funding policy is clean. Public First Action is a separate legal entity. Anthropic's interests are broadly aligned with AI safety.
• Evidence against: No disclosed recusal policy. If Anthropic's commercial interests diverge from public safety interests (e.g., on licensing requirements that would disadvantage smaller competitors), Carson faces an unarticulated conflict. The Anthropic defense in the Pentagon dispute may be principled but the financial context is relevant.
• If wrong: ARI becomes perceived as Anthropic-adjacent rather than independent, undermining its core "anti-regulatory-capture" pitch.

Strengths

Real legislative accomplishments. For a 2-year-old org with ~$5M budget, defeating a moratorium backed by Big Tech's combined lobbying apparatus and helping pass the TAKE IT DOWN Act are genuine achievements. Most policy orgs in the AI safety space have no comparable wins.

Bipartisan credibility. Carson (former D congressman) + Stewart (former R congressman). PAC gives to both parties. Worked with Cruz on TAKE IT DOWN. This is rare in the AI safety ecosystem, which skews heavily liberal/EA.

Senior Hill access. Calidas (former Klobuchar/Manchin chief of staff), Carson (former congressman with personal relationships), team full of former Hill staff. This is the kind of institutional access that takes years to build.

Clean funding model. No corporate money, no foreign money. This makes the anti-regulatory-capture argument credible in a way that many DC advocacy groups cannot match.

Strategic sophistication. The multi-entity structure (c4, c3, PAC, Carey Committee, plus Public First Action), the near-term-to-long-term causal argument, the "regulatory muscle" framing -- this is a professionally designed influence operation, not a naive advocacy effort.

"Big tent" that actually works. ARI co-leads coalitions with ITI (industry trade group), FAS (science policy), Public Citizen (consumer advocacy), and FIRE (free speech). Building coalitions across these ideological lines is genuinely hard.

Weaknesses and Risks

Funder concentration. Open Phil is ~57% of revenue. If Open Phil shifts priorities, ARI faces an existential funding crisis. The no-corporate-funding policy limits diversification options.

The Anthropic nexus is a reputational time bomb. Carson's dual role and Anthropic's $20M to Public First Action create a perception vulnerability. One well-timed investigative article connecting these dots could undermine ARI's independence narrative. The lack of a formal recusal policy makes this worse.

No technical credibility. Zero original research, zero peer-reviewed publications, zero engagement with the technical AI safety community (LW, EA Forum, AF). ARI translates technical ideas for policy audiences but does not produce them. If a policy question requires deep technical judgment (e.g., what constitutes adequate safety evaluation), ARI must rely entirely on advisors and external expertise.

The spending asymmetry is structural. Even with Public First Action, the pro-safety side is outgunned. And Big Tech's advantage is not just money -- it is revolving-door hiring, deep technical expertise, and persistent relationships. ARI's current wins came on issues where public opinion was overwhelmingly on their side. More contested issues will be harder.

Near-term wins may not transfer. The TAKE IT DOWN Act and moratorium defeat are politically popular, low-complexity wins. Frontier AI governance (compute monitoring, training run oversight, mandatory safety cases) is politically contentious, technically complex, and opposed by powerful interests. The regulatory muscle metaphor may be more metaphor than mechanism.

The "navigate uncertainty" principle could become a permanent dodge. If ARI never takes a position on whether catastrophic AI risk is real and urgent, they may never build the political will for the strongest governance measures. Bipartisan centrism can become an end in itself rather than a means to safety.

Cross-References

Compared to IAPS/AIPI: ARI is more focused on lobbying and less on research than IAPS. More established than AIPI. ARI has the strongest Hill relationships of any EA-adjacent AI policy org.

Compared to CAIP: CAIP collapsed; ARI survived and grew. ARI's moderate positioning avoided the backlash that sank CAIP's more extreme proposals (RAAIA).

Compared to Anthropic: ARI's relationship with Anthropic is the most significant cross-reference. Anthropic builds AI and funds safety research; ARI advocates for regulatory frameworks. The relationship is complementary but the financial entanglement (via Public First Action) creates dependency.

Compared to PauseAI: ARI occupies the opposite end of the AI safety advocacy spectrum. PauseAI demands a moratorium; ARI explicitly avoids taking that position. ARI's approach is "build institutions," PauseAI's is "stop the thing." Both theories could be wrong.

Compared to FAS/RAND: ARI's coalition work with FAS and its NIST/AISI advocacy overlap with government-capacity organizations. FAS does policy research; ARI does lobbying with research support.

What Would Change This Assessment

• ARI takes a specific position on catastrophic risk. If they published a statement assessing the probability or timeline of transformative AI, that would signal a strategic evolution from "navigate uncertainty" to "address specific risks."
• Anthropic's interests diverge from safety. If Anthropic takes an action that conflicts with AI safety (e.g., releasing a model without adequate testing, opposing a safety regulation that would apply to Claude), and ARI fails to criticize them or actively defends them, that would confirm the conflict-of-interest concern.
• ARI passes frontier governance legislation. If ARI helps pass compute monitoring, training run oversight, or mandatory safety evaluation requirements, that would demonstrate the "regulatory muscle transfers" hypothesis.
• Spending asymmetry collapses. If Big Tech PACs dramatically outspend Public First Action in 2026 and pro-regulation candidates lose, ARI's electoral strategy fails.
• Open Phil defunds ARI. Would test whether ARI can survive without its majority funder.

Self-Critique

Weakest claim: The assessment that ARI's near-term wins "may not transfer" to frontier governance is speculative. ARI has only existed for 2 years -- it is too early to judge whether regulatory muscle built on TAKE IT DOWN and moratorium fights will eventually be deployed on harder issues. The historical precedent of incremental regulatory development (FDA, FAA) actually supports ARI's approach.

Potential bias: This analysis may overweight the Anthropic conflict-of-interest concern. ARI's defense of Anthropic in the Pentagon dispute was widely shared across the national security community, not just by Anthropic allies. The financial connection is real but the specific action may not demonstrate captured behavior.

Missing sources: No Axios launch article (paywalled). No OpenSecrets detailed data (403 errors). No CRI financial data. No long-form interview with Gastfriend specifically about AI risk. The analysis of Gastfriend's views is thin.

What a thoughtful disagreer would say: "ARI is doing exactly the right thing. The AI safety community's obsession with frontier governance ignores the political reality that you have to build a legislative track record before anyone will listen to you on the hard stuff. ARI's bipartisan wins are the foundation for future frontier governance -- they are playing the long game while critics demand they play the end game first."

Single weakest point: The assessment lacks a strong empirical basis for evaluating ARI's counterfactual impact. Would the TAKE IT DOWN Act have passed without ARI? Would the moratorium have been defeated anyway? Attribution in coalition advocacy is inherently fuzzy, and this analysis cannot rigorously answer the counterfactual.