AI Now Institute

Stated Theory of Change

AI Now's theory of change has a clear causal chain with specific mechanisms:

Diagnosis: AI is not an autonomous technological force but a product of corporate power concentration built on the surveillance advertising business model. The technology primarily serves those who already hold power.

Intervention: Document present harms, build counter-narratives challenging inevitability claims, place staff in regulatory bodies, drive legislation through testimony and coalition work, support worker organizing and community resistance.

Mechanism: Use existing regulatory tools -- antitrust law, data privacy legislation, algorithmic accountability frameworks -- to constrain corporate AI power. "Data privacy regulation is AI regulation."

End state: A regulatory environment where companies must demonstrate safety before deployment (burden reversal), where data collection is minimized by law, and where structural competition reforms break Big Tech's AI monopoly.

This is one of the more specific and actionable theories of change in the AI policy space. It names concrete levers (antitrust, data privacy, burden-of-proof reversal) rather than just aspirational goals.

Revealed Theory of Change

Actions largely align with stated theory, with some notable additions:

1. FTC placement was the highest-impact action. Three staffers at a major regulatory agency is extraordinary for a 21-person nonprofit. This suggests the real theory of change is "staff the regulatory state with aligned personnel" -- more direct than the published reports alone would suggest.

2. International engagement is expanding faster than domestic. Kak at the UK AI Safety Summit, UN General Assembly, German Green Party, India AI Summit. This suggests AI Now may be hedging against US deregulatory trends by building international policy influence.

3. Local organizing is a new front. Data center resistance, city council testimony, community partnerships. This bottom-up approach complements the top-down regulatory strategy and may be more resilient to federal political shifts.

4. Whittaker's attention is primarily at Signal, not AI Now. Her role as Signal president ($50M/year budget) dwarfs her advisory role at AI Now. The intellectual leadership may be shifting from Whittaker to Kak/West.

5. The scope expansion -- energy, nuclear, weapons -- follows where AI power physically concentrates. This is intellectually coherent with the "AI is about power" thesis: if power flows through data centers and military contracts, that's where you intervene.

Where stated and revealed diverge: AI Now frames itself as challenging the AI industry broadly, but its actual interventions target Big Tech specifically (not, for example, Chinese AI companies, open-source AI communities, or government AI procurement). This makes sense given its US policy focus but narrows the theory of change.

Key Assumptions

1. AI risk is primarily a governance problem, not a technical one.

• Evidence for: Historical precedent (nuclear, biotech) shows governance matters; present harms from deployed AI systems are well-documented.
• Evidence against: If advanced AI systems develop genuinely novel dangerous capabilities, governance alone may be insufficient -- you need technical safety measures too.
• Testable: If catastrophic AI incidents occur that were foreseeable through technical analysis but not governance analysis, this assumption fails.
• If wrong: AI Now's entire approach addresses the wrong problem.

2. The AI bubble will burst.

• Evidence for: OpenAI losing $5B/year, Anthropic burning $5.6B, no proven profit models, historical precedent of tech bubbles.
• Evidence against: AI capabilities continue improving; enterprise adoption is growing; infrastructure investment may prove justified like early internet infrastructure.
• Testable: Observable within 2-3 years.
• If wrong: AI Now's narrative about AI being "overhyped" loses credibility, but their concerns about corporate power concentration remain valid.

3. Antitrust and data privacy regulation can meaningfully constrain AI development.

• Evidence for: EU DMA, Google antitrust rulings, growing bipartisan concern about Big Tech power.
• Evidence against: Regulatory capture, lobbying power of tech companies ($100M+/year), regulatory arbitrage to friendlier jurisdictions, political shifts toward deregulation.
• Testable: Observable through actual enforcement outcomes.
• If wrong: AI Now's primary policy tools are blunted. They would need alternative approaches.

4. Present harms and existential risks are zero-sum for policy attention.

• Evidence for: "Regulators have limited bandwidth and cannot consider all potential risks at once" (Noema article). In practice, Congressional hearings and regulatory actions do address one frame or the other.
• Evidence against: Some frameworks (like the EU AI Act) address both present harms and advanced AI risks. It's possible to have tiered regulatory approaches.
• Testable: Whether policy frameworks can effectively address both.
• If wrong: AI Now's insistence on choosing present harms OVER x-risk is unnecessarily exclusionary.

Strengths

1. Specificity of policy proposals. Unlike many orgs that offer vague calls for "responsible AI," AI Now names concrete mechanisms: data minimization, antitrust structural remedies, burden reversal. This is actionable.

2. Track record of policy access. Three staff at FTC, multiple congressional testimonies, UK AI Safety Summit representation, UN General Assembly remarks. For a 21-person org, this is exceptional reach.

3. No corporate funding. This is genuinely distinctive and makes their criticism of AI companies more credible than orgs that accept Big Tech money while criticizing Big Tech practices.

4. Coherent intellectual framework. Whether you agree or not, the "AI is corporate power concentration" thesis is internally consistent and produces clear predictions. This makes it falsifiable, which is a strength.

5. Expanding scope tracks reality. Moving into energy policy, nuclear regulation, and weapons systems as AI physically concentrates in those domains shows intellectual agility.

6. Lean operation. ~21 people producing this level of policy influence suggests high efficiency.

Weaknesses and Risks

1. Categorical dismissal of x-risk is intellectually vulnerable. Saying x-risk is "ghost stories" with "no basis in evidence" is a stronger claim than the evidence supports. A more defensible position would be "x-risk is speculative and should not displace attention to documented present harms" -- which is different from "x-risk is essentially unfounded." If AI capabilities continue advancing, this dismissal becomes a liability.

2. Financial opacity undermines credibility. An org that advocates for transparency in AI systems has no public budget, no visible board, and no 990 filings accessible. This creates an easy target for critics and limits independent assessment.

3. Funder concentration. Omidyar ecosystem provides multiple current funders. The org is dependent on the continued interest of mainstream progressive philanthropy. If foundation priorities shift (as they regularly do), AI Now has no fallback revenue.

4. Policy theory depends on political environment. AI Now's theory of change worked best during the Biden administration (FTC placements, sympathetic regulators). In a deregulatory environment, their primary levers are weakened. Their 2025 report acknowledges this but doesn't articulate a clear alternative strategy.

5. No engagement with technical AI safety. By ignoring alignment research, interpretability, and formal verification entirely, AI Now has no position on whether these technical approaches are valuable. If technical safety measures prove necessary for managing advanced AI, AI Now will have contributed nothing to that effort.

6. Single narrative vulnerability. The "AI is a bubble" claim is high-risk. If AI finds viable business models and continues improving, AI Now's core narrative loses credibility even if their concerns about corporate power remain valid.

Cross-References

• Complementary to DAIR (Timnit Gebru): Same intellectual tradition, both focus on present harms, partner organizations. DAIR is more focused on specific technical biases; AI Now is more focused on structural/policy interventions.
• Opposed to MIRI, ARC, METR: AI Now explicitly rejects the x-risk framing that drives these organizations. Zero engagement between the ecosystems.
• Overlapping with Partnership on AI, Ada Lovelace Institute, Data & Society: All in the "responsible AI" space but AI Now is more explicitly anti-corporate and politically engaged.
• Tension with OpenAI, Anthropic safety teams: AI Now views safety commitments from AI labs as "grading their own homework." They don't acknowledge that some lab safety researchers may share their concerns about corporate incentives.
• Signal Foundation as organizational sibling: Deep leadership overlap (Whittaker, Kak). Signal's privacy-first model is the practical embodiment of AI Now's theoretical critique of surveillance capitalism.
• Potential bridge figure: Heidy Khlaaf. Her traditional safety engineering background (nuclear, autonomous vehicles) and her work evaluating AI weapons systems could connect AI Now to the broader AI safety community if the org chose to engage.

What Would Change This Assessment

• AI Now publishes 990 or discloses budget: Would enable proper financial analysis and change confidence on sustainability.
• Specific policy attribution documented: If a specific law or regulation can be traced to AI Now's work, that dramatically strengthens the theory of change assessment.
• AI bubble bursts: Would validate AI Now's central empirical claim and strengthen their credibility.
• AI capabilities plateau: Would validate Whittaker's claims about fundamental technical limitations.
• Catastrophic AI incident from governance failure vs. technical failure: If a major AI harm occurs that governance alone couldn't have prevented (but technical safety measures could have), it would expose the limitation of AI Now's governance-only approach.
• AI Now engages with technical AI safety community: Would signal intellectual expansion and could produce the most productive disagreements in the field.
• Board composition revealed: Would enable governance quality assessment.

Self-Critique

Weakest claims:

• My assessment of AI Now's budget ($3-6M range) is speculative inference from team size and known grants, not verified data.
• The "Crawford departure is significant" claim may overweight a normal professional transition.

Potential biases:

• This analysis may be too sympathetic to the x-risk counterargument because the research project is embedded in an ecosystem (this company analysis tool) that takes catastrophic AI risk seriously. An analyst from AI Now's world would likely find my treatment of their x-risk dismissal unfairly critical.
• The emphasis on financial opacity may be disproportionate given that many similar-sized nonprofits have limited public financial data.

What a thoughtful disagreer would say:

• "You're treating AI Now's rejection of x-risk as a weakness, but their whole point is that x-risk discourse actively harms the people they're trying to protect. Asking them to engage with x-risk is like asking a climate scientist to engage with climate denial -- it legitimizes a framework they consider harmful."
• "The financial opacity critique is valid but you give it too much weight relative to the org's actual policy output and impact."

Single weakest claim: My suggestion that AI Now's "bubble will burst" prediction is high-risk. They may be right, and even if wrong about timing, the underlying concern about AI companies' business model sustainability is well-grounded in financial analysis.

Information that would most change my view: Seeing AI Now's actual budget, board composition, and any internal documentation of policy wins. Also, any private communication between AI Now and x-risk researchers -- if engagement happens behind closed doors that would change the "zero engagement" assessment significantly.