Partnership on AI

Stated Theory of Change

PAI's stated theory of change is: multi-stakeholder convening produces collectively agreed guidelines, which change industry practice and inform government policy, resulting in AI that benefits people and society.

The specific mechanism, from PAI's 2021 Strategic Plan: diverse stakeholders (industry, civil society, academia) convene around AI challenges -> produce research, tools, and guidelines -> partners adopt guidelines, policymakers reference them -> practice changes and policy improves.

This is a standard theory of change for a convening/standards body. It works in other domains (e.g., ISO standards, the IPCC). The critical question is whether it works when the entities whose behavior needs to change are also the primary funders and governors of the convening body.

Revealed Theory of Change

PAI's actions reveal a different emphasis than its stated theory suggests:

What PAI actually optimizes for is institutional relevance and partner satisfaction. Its outputs consistently land in the zone of "helpful guidance for well-intentioned actors" rather than "enforceable standards that constrain harmful actors." The Synthetic Media Framework is "like a constitution, not a set of laws." The Model Deployment Guidance provides 22 recommendations but no accountability mechanism. PAI explicitly states it "will not be auditing or certifying organizations."

Evidence that PAI's revealed priorities diverge from its stated mission:

1. The risk assessment report (2019): Internal documents show researchers wanted to declare algorithmic risk tools should not be used in pretrial detention. The final version was softened to say the tools are "unfit for use" but offered guidance for "minimum requirements for responsible deployment" -- a framework that critics said "validates the use of risk assessment by emphasizing the issue as a technical one."

2. Corporate contributions are non-earmarked by design, but the structural dependency is clear: PAI's revenue is ~90%+ contributions, its board chair is always a Big Tech executive, and it has never taken a position that would seriously threaten any funder's business model.

3. PAI positions itself as a coordination body between governments and industry, not as an advocate for specific regulatory outcomes. Its 2026 priorities document discusses "interoperability" between governance frameworks -- a framing that implies industry needs harmonized rules, not stricter ones.

The revealed theory of change is closer to: PAI provides a venue where companies can demonstrate good faith engagement with AI governance, civil society can participate in discussions they otherwise wouldn't access, and policymakers can reference PAI's outputs as evidence of industry self-governance, reducing pressure for binding regulation.

This isn't necessarily bad. But it's different from what PAI claims.

Key Assumptions

Assumption 1: Multi-stakeholder bodies can produce outputs that constrain their most powerful members.

• Evidence against: The Intercept's documentation of how corporate interests shaped PAI's criminal justice recommendations. Access Now's finding that PAI "had not influenced or changed the attitude of member companies." Academic literature on corporate capture of multi-stakeholder initiatives.
• Evidence for: The risk assessment report did conclude tools were "unfit for use" -- a position contrary to the tech industry's general preference for technical fixes over bans. The ACLU maintained 7+ years of board engagement, suggesting some civil society actors find value.
• Testable: If PAI ever produces a recommendation that a founding member publicly objects to or acts against, this assumption gains credibility.
• If wrong: PAI is an expensive talking shop that creates the appearance of governance without the substance. Industry gets a legitimacy shield while civil society wastes resources on a process that cannot produce meaningful change.

Assumption 2: Voluntary guidelines, widely adopted, can substitute for or complement regulation.

• Evidence against: Only 2 of 128 partners are publicly documented as implementing PAI guidelines. No compliance verification exists. The academic study found 14 of 24 "committed" firms took no concrete implementation steps.
• Evidence for: PAI's work has been cited by the FTC, NIST, OECD, and the White House AI Bill of Rights. The EU AI Act references transparency requirements that PAI helped develop. This is evidence of guidelines informing regulation, which is the complement case.
• If wrong: Voluntary guidelines become "ethics washing" -- performative commitment that delays binding action.

Assumption 3: Near-term AI harms (bias, fairness, media integrity) are the primary governance challenge.

• Evidence for: Near-term harms are real, documented, and happening now. AIID cataloged 1,200+ real incidents.
• Evidence against: If advanced AI poses existential or catastrophic risks, governance focused exclusively on near-term deployment harms fails to address the most important scenarios. PAI's entire framework assumes AI systems work roughly as intended and asks how to govern their deployment.
• If wrong: PAI is addressing the wrong problem at a time when the most important governance questions are about capabilities, not deployment.

Strengths

1. Early mover with institutional staying power. Founded 2016, still operational with 128 partners and $6-9M annual revenue. Most AI ethics initiatives from this era have faded. PAI has institutional continuity.

2. Genuine policy influence. Work cited by the White House, FTC, NIST, OECD, EU AI Office. The ABOUT ML project was specifically referenced in the AI Bill of Rights. PAI has demonstrated ability to get its outputs into policymakers' hands.

3. AI Incident Database. A genuine public good that spun off to independence. The AIID model -- systematically cataloging AI failures -- has real value for the field regardless of PAI's other limitations.

4. Synthetic Media Framework. 18 institutional supporters with 16 published case studies is meaningful adoption. This is PAI's best example of turning convening into practice change.

5. Sustained civil society engagement. The ACLU maintained board-level involvement for 7+ years. Ford Foundation, Brookings, universities, and international organizations participate. This is not a purely industry body.

6. Workforce focus. PAI's Shared Prosperity Guidelines and data enrichment work address an issue (AI's impact on workers) that gets too little attention in the AI safety community.

Weaknesses and Risks

1. Structural conflict of interest is unresolved. The companies whose behavior PAI aims to influence fund the organization, govern its board, and face no consequences for ignoring its recommendations. The Intercept article documents what this looks like in practice. Access Now's resignation confirms the structural problem from the civil society perspective. Eight years later, the board chair is still a Big Tech executive.

2. No enforcement mechanism. PAI explicitly declines to audit or certify. Guidelines are voluntary. Only 2 of 128 partners are publicly documented as implementing PAI's recommendations. Without enforcement, the gap between PAI's output and actual practice change remains unmeasurable.

3. Funding opacity. For an organization that advocates AI transparency, PAI does not disclose individual corporate contribution amounts, staff headcount, partner attrition data, or the reasons for its 2024 executive compensation spike. This undermines credibility.

4. No engagement with the most important AI governance questions. If advanced AI poses catastrophic or existential risks, PAI's exclusive focus on near-term deployment harms means it is working on secondary problems. PAI has never taken a position on capability restrictions, moratoriums, or compute governance.

5. The "legitimate but captured" pattern. PAI provides a venue for civil society participation and produces research that policymakers cite. This creates a self-reinforcing legitimacy loop where PAI's existence reduces pressure for stronger governance. Even if individual outputs are high quality, the institutional effect may be to channel governance energy into a framework designed not to threaten incumbent interests.

6. Declining relevance. The Frontier Model Forum (OpenAI, Google, Microsoft, Anthropic) now handles frontier AI safety coordination directly. Government AI Safety Institutes have been established in the US, UK, and other countries. The EU AI Act provides binding regulation. Each development reduces the governance space where PAI operates.

Cross-References

• GovAI (Centre for the Governance of AI): GovAI's Markus Anderljung participated in PAI's Model Deployment Guidance working group. GovAI does policy research that PAI's outputs sometimes draw on, but GovAI focuses on advanced AI governance including existential risk -- a domain PAI avoids.

• Frontier Model Forum: Separate from PAI, industry-only (OpenAI, Google, Microsoft, Anthropic). The FMF "explicitly coordinates with PAI" but handles frontier safety commitments directly. PAI's role may be increasingly limited to near-term governance.

• AI Now Institute: PAI partner. Founded by researchers with Microsoft connections. AI Now takes stronger advocacy positions than PAI. Their coexistence shows the spectrum from "convening" (PAI) to "advocacy" (AI Now).

• AISI (UK/US AI Safety Institutes): Government bodies that now do safety evaluation work that PAI was positioned to influence but never conducted. Their establishment reduces PAI's governance relevance.

• METR, ARC Evals: Technical AI evaluation organizations in the AI safety ecosystem. Doing the kind of concrete safety work (evaluating model capabilities) that PAI's framework assumes someone else will do.

What Would Change This Assessment

• PAI publishes a recommendation that a founding member publicly objects to or acts against. This would demonstrate genuine independence and would be the strongest possible evidence against the "captured" thesis.

• PAI begins work on advanced AI risk, compute governance, or capability restrictions. This would show PAI can evolve beyond its near-term harm focus.

• Independent evaluation of PAI's impact on partner behavior. Self-reported annual reports are insufficient. A rigorous external evaluation showing measurable practice changes would significantly strengthen PAI's case.

• Disclosure of corporate contribution amounts. Full funding transparency would allow external evaluation of whether PAI's positions correlate with funder interests.

• Multiple civil society organizations publicly depart PAI with statements similar to Access Now's. This would confirm the structural critique is representative, not exceptional.

Self-Critique

Weakest claim: My characterization of PAI as structurally captured relies heavily on The Intercept's 2019 article and Access Now's 2020 resignation. Both are dated. PAI has refreshed its board (2024), grown to 128 partners, and produced more substantive work (2023-2024). It's possible the dynamics documented in 2019 have improved. The ACLU's sustained engagement suggests some civil society organizations find genuine value.

Potential bias: I may overweight the "ethics washing" critique because it's a dramatic narrative and the evidence (The Intercept) is vivid and specific. PAI's day-to-day work producing guidelines, hosting forums, and informing policymakers is less narratively compelling but may be more impactful than the dramatic critique suggests.

What I should have checked: I lack access to PAI's Schedule B filings (donor amounts over $5K), which would reveal the degree of corporate funding concentration. I also lack data on what happens inside PAI's working groups -- whether civil society voices genuinely influence outputs or are marginalized. This kind of process evidence is not publicly available.

Single weakest claim: My statement that PAI's "revealed theory of change" is about providing legitimacy cover for industry. This is an inference from structural incentives and limited evidence of captured outputs (mainly the 2019 risk assessment episode). It's possible PAI's outputs are genuinely independent and the structural conflicts are mitigated by the governance safeguards PAI describes.

What would most change my view: A detailed, independently conducted study showing that PAI partner organizations changed specific practices as a direct result of PAI guidelines, compared to non-partner organizations. This would demonstrate that PAI's convening model actually drives practice change, which is the core claim in its theory of change.