Center for AI and Digital Policy (CAIDP)

Stated Theory of Change

CAIDP believes that AI can be governed through democratic institutions and international legal frameworks. Their causal chain:

1. Establish global norms (OECD AI Principles, UNESCO Recommendation, Universal Guidelines for AI) that define what responsible AI governance looks like.
2. Measure compliance by scoring countries against these norms annually (AIDV Index), creating reputational incentives for improvement.
3. Enforce accountability by filing legal complaints (FTC) and amicus briefs when companies or governments violate existing law.
4. Build capacity by training practitioners in 120+ countries who will staff the next generation of AI regulatory bodies.
5. Promote binding treaties (CoE Framework Convention on AI) that create legal obligations enforceable across borders.

The underlying assumption is that democratic governance, human rights law, and international cooperation can move fast enough and scale broadly enough to constrain AI development within safe bounds.

Revealed Theory of Change

CAIDP's actions are broadly consistent with its stated theory. The AIDV Index is genuinely produced. The FTC complaint was genuinely filed and prompted an investigation. The policy clinics genuinely train people. The CoE treaty is genuinely signed by 41 countries.

However, several subtle divergences emerge:

The Index measures policy adoption, not policy effectiveness. The AIDV scores countries on whether they endorse OECD principles, not on whether their AI systems actually respect human rights. Canada ranking #1 does not mean Canadian AI is safer than American AI -- it means Canada has more formal governance structures. This is scoring the menu, not tasting the food.

The legal strategy aims for attention rather than relief. The FTC complaint asked to halt GPT-4 deployment -- a remedy the FTC almost certainly lacked authority to grant. The actual goal was to force the FTC to engage with AI oversight, which succeeded. This is advocacy masked as litigation.

The global network is volunteer-driven. 1,000+ contributors in 120 countries sounds like a massive operation, but these are unpaid policy clinic participants producing country reports. The research depth per country is necessarily limited. This is breadth traded for depth.

International frameworks are aspirational, not binding. The OECD AI Principles, UNESCO Recommendation, and even the CoE AI Treaty lack enforcement mechanisms against the actors most likely to create dangerous AI systems (primarily US and Chinese tech companies). The frameworks create diplomatic language, not safety constraints.

Key Assumptions

Assumption 1: Democratic governance can move fast enough.

• For: The EU AI Act, CoE treaty, and 120+ country UN AI resolution show governments can act on AI governance.
• Against: These frameworks took 3-5 years to develop. AI capabilities are advancing on 6-12 month timescales. The gap between governance speed and technological speed may be widening.
• Testable: Whether any AI-specific regulation prevents a demonstrable harm within 2 years of passage.
• If wrong: Democratic governance becomes permanently reactive rather than preventive.

Assumption 2: Existing legal frameworks (consumer protection, human rights) are adequate tools for AI governance.

• For: Section 5 of the FTC Act prohibiting "unfair and deceptive practices" is broad enough to cover many AI harms. The CoE treaty explicitly addresses AI.
• Against: These frameworks were designed for human decision-makers and existing technologies. Novel risks from advanced AI (power-seeking behavior, loss of control, deceptive alignment) may require entirely new legal concepts.
• Testable: Whether existing consumer protection law can address a catastrophic AI failure.
• If wrong: CAIDP's entire approach needs supplementation with technical safety mechanisms it currently ignores.

Assumption 3: Country scoring creates meaningful incentive pressure.

• For: Countries do care about international rankings (see OECD PISA scores' impact on education policy).
• Against: No evidence that any country changed its AI policy in response to an AIDV ranking. The countries most likely to deploy dangerous AI (US, China) are least susceptible to scoring pressure.
• Testable: Whether countries that score poorly in the AIDV subsequently improve.
• If wrong: The Index is an academic exercise rather than a policy tool.

Assumption 4: Training 1,500+ practitioners across 120 countries builds meaningful governance capacity.

• For: Many countries lack any AI policy expertise. Even basic training fills a real gap.
• Against: Semester-long clinics producing country reports may not develop the deep expertise needed to regulate frontier AI systems. The alumni network's actual policy influence is unmeasured.
• Testable: Whether clinic alumni go on to hold positions that influence AI governance.
• If wrong: The clinics are valuable as education but not as capacity-building for effective governance.

Strengths

Genuine independence. Zero funding from AI labs, tech companies, or the x-risk philanthropy ecosystem. CAIDP can criticize any company or government without financial conflict. This is rare and valuable.

Extraordinary leverage per dollar. On approximately $924K revenue (2024), CAIDP produces an 80-country index with 1,000+ contributors, maintains 90,000 newsletter subscribers, prompted an FTC investigation, and trains hundreds of practitioners annually. The output-to-budget ratio is remarkable.

Bridging position. By signing the Pause letter while being rooted in the rights/ethics tradition, CAIDP occupies a unique position connecting two communities that usually talk past each other. Rotenberg's explicit acknowledgment that the Universal Guidelines' "termination obligation" maps onto existential risk concerns shows intellectual flexibility.

Rotenberg's network and credibility. 40+ years of technology policy work, 60+ Congressional testimonies, OECD Expert Group membership, Georgetown Law faculty position. This accumulated social capital is CAIDP's primary asset and is essentially irreplaceable.

Concrete track record. The FTC investigation of OpenAI is a real, traceable policy outcome. The CoE AI Treaty, while CAIDP's specific contribution is hard to isolate, is a real international agreement that CAIDP advocated for.

Global South inclusion. 64% of countries with meaningful AI policies in the AIDV are in Africa, Asia, Latin America, Caribbean, Middle East, and Oceania. CAIDP's clinic deliberately recruits from underrepresented regions. Most AI governance organizations are Western-centric; CAIDP's global network is genuinely unusual.

Weaknesses and Risks

May be solving the wrong problem. If the primary AI risk is catastrophic/existential (loss of control, deceptive alignment, power-seeking), then CAIDP's approach -- which addresses democratic governance, bias, transparency, and human rights -- may be structurally inadequate. Human rights frameworks were designed for present harms to identifiable individuals, not for tail-risk scenarios affecting all humans simultaneously. CAIDP has zero engagement with technical AI safety.

Key person risk is extreme. CAIDP is essentially Marc Rotenberg. The funding relationships, institutional networks, and policy expertise are almost entirely personal. Rotenberg is 65. There is no visible succession plan. If he is unable to lead, the organization may not survive in any meaningful form.

Governance is weak. A three-person board where two members are the primary staff means no real independent oversight. For an organization that advocates governance standards for AI, the irony of lacking governance standards for itself is sharp.

Impact measurement is absent. CAIDP measures 80 countries but has not demonstrated that its own work changes any country's policy. The FTC investigation is the strongest evidence, but one action does not validate the entire theory of change. Whether the AIDV Index or policy clinics have traceable outcomes remains unproven.

Funder concentration. The McGovern Foundation's $500K grant may represent over half of CAIDP's revenue, creating significant dependency on a single funder.

Speed mismatch. Annual index publications, multi-year treaty negotiations, and semester-long clinics operate on timescales that may be too slow for AI governance. If frontier AI capabilities advance dramatically within 1-2 years, CAIDP's approach may become permanently behind.

Civil society exclusion. The ECNL warning that states excluded civil society from CoE treaty drafting suggests that CAIDP's "inside" advocacy approach has real limits. Being an observer is not the same as having influence.

Cross-References

Complementary to GovAI/CLTR: CAIDP and x-risk governance organizations (GovAI, CLTR) are working on the same broad problem from different intellectual traditions and with different tools. CAIDP brings international legal frameworks and developing-world inclusion; GovAI/CLTR bring technical understanding of AI risk and frontier lab engagement. Zero current collaboration exists, which represents a missed opportunity.

Different from PauseAI/ControlAI: While CAIDP signed the Pause letter, its approach is fundamentally institutional and incremental -- filing FTC complaints, building treaties, training practitioners. PauseAI and ControlAI seek more radical interventions. CAIDP provides the establishment governance track to their activist track.

Gap-filling for Stanford HAI: Stanford HAI has more resources and academic credibility but is US-focused and has significant industry ties. CAIDP's genuine independence from AI labs and its Global South network fill gaps that HAI cannot.

Parallel to CDT, Access Now, EFF: CAIDP operates in the same digital rights ecosystem as these organizations but is more narrowly focused on AI governance specifically. Its Index product differentiates it from peers that do broader tech policy.

What Would Change This Assessment

• Upward: Evidence that an AIDV ranking caused a specific country to change its AI policy. Evidence that CAIDP policy clinic alumni are now in positions shaping AI governance. A second major regulatory action traceable to CAIDP advocacy. Collaboration with technical AI safety organizations.

• Downward: Loss of McGovern Foundation funding without replacement. Rotenberg's departure without a credible successor. Evidence that the AIDV methodology systematically overstates governance quality (scoring policy adoption rather than implementation). An AI incident that democratic governance frameworks prove completely inadequate to address.

• Sideways: If AI risk turns out to be primarily about present harms (bias, surveillance, labor displacement) rather than catastrophic/existential risk, CAIDP's approach becomes much more central to the solution. The assessment depends heavily on your prior about which risks matter most.

Self-Critique

What sources should I have checked but didn't? The 2022 annual report PDF (raw binary), which would contain board members and program details. The Taylor & Francis academic critique of AIDV methodology (paywalled). Rotenberg's forthcoming casebook (The Law of AI) for his detailed legal framework. CAIDP's actual about/team pages (blocked by 403).

Where is this analysis potentially biased? I approach from the x-risk/alignment perspective, which may cause me to undervalue the importance of democratic governance and human rights frameworks. CAIDP's theory of change may be more powerful than I credit if AI risk turns out to be primarily about present-day harms rather than existential threats. I may also be too generous about the FTC complaint's impact -- the investigation may have happened independently of CAIDP's filing.

What would a thoughtful person who disagrees say? "You are evaluating CAIDP through an x-risk lens when CAIDP operates in a different frame. The OECD Principles and CoE Treaty are binding international law -- not aspirational statements -- and they apply to all AI systems, including frontier systems. Democratic governance of technology has a far longer and more successful track record than technical alignment research, which has produced no verifiable safety guarantees. CAIDP's approach of building global consensus and institutional capacity is the only approach that scales."

What is my single weakest claim? That CAIDP's approach "may be solving the wrong problem." If catastrophic AI risk manifests primarily through misuse (autonomous weapons, mass surveillance, targeted disinformation) rather than through loss-of-control scenarios, then human rights frameworks are precisely the right tool.

What information would most change my view? Evidence that a country changed its AI development practices in response to an AIDV ranking. Evidence that CAIDP's policy clinic alumni have influenced concrete AI governance outcomes. Evidence that Rotenberg has engaged seriously with x-risk arguments and has a considered response, rather than simply operating in a separate intellectual tradition.