Partnership on AI

Theory of Change

PAI's stated theory of change: multi-stakeholder convening produces guidelines and resources, which inform practice change and policy innovation. CEO Rebecca Finlay: "We develop tools, recommendations, and other resources by inviting voices from across the AI community and beyond to share insights that can be synthesized into actionable guidance. We then work to drive adoption in practice, inform public policy, and advance public understanding."

PAI's 2021 Strategic Plan identifies four intended outcomes: (1) an inclusive PAI community, (2) a better-informed public, (3) policy innovations by governments, (4) changes in practice by partners. The mechanism is convening -> research -> guidelines -> adoption.

PAI explicitly states it is "not a trade group or advocacy organization" and "will not be auditing or certifying organizations." The Synthetic Media Framework is described as "like a constitution, not a set of laws." There is no enforcement mechanism for any PAI output.

PAI has zero engagement with existential risk, AI alignment, or catastrophic AI scenarios. Its entire framework addresses near-term deployment harms: bias, fairness, media integrity, worker treatment, governance coordination.

What They Do

Most significant outputs:

AI Incident Database (2020-2022): Cataloged 1,200+ real-world AI failures. Spun off to independent nonprofit in 2022. Arguably PAI's most lasting concrete contribution -- a genuine public good.
Synthetic Media Framework (2023-present): 18 institutional supporters (BBC, Google, Meta, TikTok, Adobe, OpenAI). 16 case studies published in 2024. PAI's most widely adopted output and strongest evidence of engagement.
Model Deployment Guidance (2023): 22 guidelines for foundation model providers, tiered by capability and release type. Developed with 50+ experts including GovAI's Markus Anderljung.
Risk Assessment Tools Report (2019): Concluded algorithmic risk assessment tools are "unfit for use" in pretrial detention. PAI's most substantive policy stance. Internal documents reveal this conclusion was a compromise -- some researchers wanted stronger language.
Data Enrichment Sourcing Guidelines: Only two companies (DeepMind, OpenAI) are publicly documented as adopters.
ABOUT ML documentation project: Cited in the White House AI Bill of Rights (2022). Three pilot studies published.
Open Model Value Chain mapping (2024): Technically detailed analysis of risk mitigation strategies across the AI value chain, co-hosted with GitHub.

Policy influence markers: PAI's work has been cited by the FTC, NIST, OECD, the White House AI Bill of Rights, and referenced in EU AI Act transparency provisions. PAI hosted Policy Forums in London (2023) and New York (2024).

Notable absence: No work on compute governance, AI capability restrictions, or moratoriums. All outputs assume continued AI development and ask how to govern deployment.

Key People

Rebecca Finlay, CEO since October 2021. Former VP at CIFAR (Canadian Institute for Advanced Research) where she founded the AI & Society program. Not a technologist -- brings nonprofit management and science policy experience. Her podcast interviews are thoughtful, and she explicitly endorses regulation alongside voluntary frameworks.

Jerremy Holland (Apple, Director of AI Research) has chaired the board since October 2023, succeeding Eric Horvitz (Microsoft CSO), who was founding chair for 8 years. The board chair position has been held exclusively by Big Tech executives since founding.

Notable departures: Peter Eckersley, Director of Research (2018-2020), formerly EFF's Chief Computer Scientist, left to co-found the AI Objectives Institute (deceased 2022). Terah Lyons, founding Executive Director (2017-2021), departed to eventually become Global Head of AI Policy at JPMorgan Chase.

Team size: ~30-40 staff (estimated from salary data). Exact headcount not publicly disclosed.

Money and Incentives

Revenue: $6-9M/year. Peaked at $10.2M in 2018, stabilized around $6-8M. 2024 revenue ~$9M.

Revenue breakdown: ~90-100% contributions. No product revenue. Rental property income provides ~$400-500K/year (~6%). Revenue is entirely donation/grant-dependent.

Founding corporate funders: Apple, Amazon, Meta, Google/DeepMind, IBM, Microsoft provided "multi-year grants" at founding. Amount of each company's ongoing contributions is not publicly disclosed.

Named philanthropic funders: MacArthur Foundation ($3.3M over 2017-2025, ~$400K/year). Luminate (Omidyar Group, amount undisclosed).

Corporate partner model: Each for-profit partner makes an "annual charitable contribution" upon joining. With 128 partners, these contributions likely form a significant but undisclosed portion of revenue.

Net assets: Declining from $13.9M (2018) to $6.6M (2023), suggesting PAI has been spending down initial reserves.

Executive compensation: $2.29M in 2024, representing 27.4% of total expenses. CEO Finlay earned $337K + $31K other. The remaining ~$1.9M went to other executives. No public explanation for this spike. In normal years (2022-2023), total executive compensation was $432K-$868K (5-11% of expenses).

Incentive structure: PAI's survival depends on contributions from Big Tech companies. Board governance is shared between for-profit and nonprofit directors by bylaw, but the chair has always been a Big Tech executive. PAI's funding model explicitly classifies corporate contributions as "non-earmarked charitable contributions" to prevent formal conflicts of interest. However, the simpler mechanism -- organizational reluctance to antagonize funders -- requires no earmarking.

990 anomaly: The "contributions" line shows $31-37M for 2021-2023 versus $6-7M in actual revenue, indicating multi-year pledge accounting.

What Others Say

The strongest critique (The Intercept, 2019): Former MIT Media Lab researcher Rodrigo Ochigame reports that PAI's policy recommendations on criminal justice "aligned consistently with the corporate agenda." Joichi Ito privately told a billionaire that PAI "waters down stuff we try to say." In a private meeting with Ito, PAI co-founder Mustafa Suleyman acknowledged that PAI's promotion of "AI ethics" had become "whitewashing." An internal email stated: "Neither ACLU nor MIT nor any non-profit has any power in PAI."

Access Now resignation (2020): "We did not find that PAI influenced or changed the attitude of member companies or encouraged them to respond to or consult with civil society on a systematic basis." Access Now advocates for outright bans on certain technologies; PAI's framework/risk-assessment approach was fundamentally insufficient.

Academic analysis (2021): Study of 24 companies committed to responsible AI found 14 had taken no concrete implementation steps. Discusses PAI as a voluntary commitment with no enforcement mechanism.

Ethics washing framing: Carnegie Council defines ethics washing as "creating a superficially reassuring but illusory sense that ethical issues are being adequately addressed." Academic literature specifically cites industry-funded AI ethics initiatives as exemplars of this pattern.

Defense -- Finlay: "None of the work that we're doing at the Partnership on AI in any way should stop appropriate regulation. I've always been a supporter of governments attending to and being aware of and acting upon harms." She argues PAI serves companies that want to be responsible, gives civil society a seat at the table, and informs policymakers.

Defense -- sustained engagement: The ACLU maintained board-level engagement from founding through 2024 (7+ years). Carol Rose (ACLU) received a PAI Changemaker Award. If PAI were purely performative, sustained ACLU engagement would be hard to explain.

Community visibility: PAI has zero presence on LessWrong or EA Forum. No dedicated profile in any major tech publication (Wired, MIT Tech Review, The Verge). PAI operates below the radar of both the AI safety community and mainstream tech journalism.

What's Absent

No enforcement or compliance mechanism. PAI's guidelines are entirely voluntary.
Corporate funding amounts undisclosed. We cannot determine what each Big Tech company contributes.
No verified adoption data. Only 2 of 128 partners are publicly documented as implementing PAI guidelines.
No whistleblower or conflict-of-interest policies publicly documented.
No independent evaluation of impact has been commissioned.
No engagement with existential AI risk, alignment, or capability restrictions.
Staff headcount not disclosed despite promoting transparency.
Partner attrition not tracked publicly -- we know Access Now left but not who else.
2024 executive compensation spike (27.4% of expenses) has no public explanation.