The Midas Project

Stated Theory of Change

TMP explicitly transplants the animal welfare corporate campaigns model to AI safety. The mechanism: build a credible threat of public pressure against AI companies, making the reputational and political cost of ignoring safety concerns exceed the cost of compliance. Start small (Cognition), build reputation and supporter base, scale to larger targets (OpenAI).

Johnston describes TMP as the "bad cop" in a good-cop/bad-cop ecosystem. Expert policy groups (GovAI) and "good cop" collaborative orgs handle the inside game. Radical activists (PauseAI, Stop AI) shift the Overton window. TMP fills the middle: permanent institutional pressure on specific companies, using investigations, open letters, legal complaints, and public campaigns.

The stated goal is not to shut down AI development but to make safety commitments credible and enforceable -- tracking broken promises, exposing conflicts of interest, and creating consequences for companies that cut corners.

Revealed Theory of Change

TMP's actions reveal an evolution from the stated corporate campaigns model toward something more like investigative journalism and legal/regulatory entrepreneurship.

What started as corporate campaigns (the Cognition campaign, 2024) quickly shifted to investigative compilation (OpenAI Files, 2025), then to legal/regulatory action (IRS complaint, SB 53 violation allegation, AG outreach), and most recently to investigative journalism (Model Republic a16z report, astroturf exposures).

This evolution makes sense as an adaptation: the corporate campaigns model depends on compliance costs being bounded, and Johnston recognized early that AI compliance costs are not bounded. So TMP pivoted to approaches that do not depend on the cost calculus: legal compliance is binary (you either follow the law or you do not), governance transparency has legal requirements for nonprofits, and investigative journalism creates accountability regardless of whether the target changes behavior.

The revealed theory of change is closer to: create an information commons and legal accountability infrastructure that makes it harder for AI companies to break promises without consequence. This is more modest but also more defensible than the original corporate campaigns framing.

However, there is a significant gap between TMP's output (which is impressive) and measurable impact (which is nearly zero by documented evidence). No company has demonstrably changed behavior in response to TMP's work. The organization produces accountability journalism and files complaints, but the causal chain from these outputs to actual risk reduction remains unproven.

Key Assumptions

1. Public pressure matters for AI companies.

• Evidence for: AI companies do care about reputation (OpenAI's restructuring narrative, Anthropic's safety branding). The subpoena campaign itself is evidence that OpenAI views TMP as threatening enough to try to silence.
• Evidence against: Johnston's own analysis -- compliance costs may be unbounded, so no amount of pressure may be sufficient. Tech companies have enormous PR resources. Public attention is fleeting.
• Testable: Partially. Track whether companies change behavior after TMP campaigns.
• If wrong: TMP becomes an accountability archive rather than a change agent.

2. The "bad cop" role requires organizational durability.

• Evidence for: Johnston's Humane League experience shows this works over years, not months. The credible threat requires sustained presence.
• Evidence against: At $25K/yr salary, 1 FTE, and post-subpoena uninsurability, organizational survival is not guaranteed. The animal welfare bad cops have sustainable funding; TMP may not.
• Testable: Track whether TMP still exists and is producing work in 2-3 years.
• If wrong: TMP becomes a cautionary tale about the limits of adversarial advocacy against trillion-dollar companies.

3. Investigations and legal complaints create meaningful accountability.

• Evidence for: The SB 53 violation allegation created a real regulatory response (CA AG statement). IRS complaints can trigger investigations. Open letters create public records.
• Evidence against: No IRS action reported. No AG investigation confirmed. The legal system is slow and companies have unlimited legal resources.
• Testable: Track regulatory outcomes from TMP's complaints.
• If wrong: TMP's legal complaints become performative rather than functional.

4. Cross-movement expertise transfer works.

• Evidence for: TMP's output quality is remarkably high given its budget. Johnston's communications skills from animal welfare translate well to AI accountability.
• Evidence against: The key element that made animal welfare campaigns work (bounded compliance costs) does not transfer. The skills transfer; the strategic logic may not.
• Testable: Compare TMP's trajectory to other cross-movement organizations.
• If wrong: Animal welfare is a misleading analogy, and TMP needs a fundamentally different strategic framework.

Strengths

• Extraordinary output per dollar. One person on $25K/year producing 18K-word investigative reports, organizing open letters with Nobel laureates, building tracking tools, and filing legal complaints. The dollar-for-dollar impact on information production is probably unmatched in AI safety advocacy.

• Non-partisan targeting. TMP criticizes OpenAI, Anthropic, xAI, and a16z. It signed a letter against xAI deepfakes while simultaneously publishing the OpenAI Files. This makes the "funded by Musk" accusation look absurd and builds credibility.

• Intellectual honesty about limitations. Johnston's EA Forum comments show genuine grappling with whether his approach works, including the bounded/unbounded cost problem and the perverse incentives problem. This is rare in advocacy organizations.

• Filling a real gap. The AI safety ecosystem genuinely needed an adversarial accountability organization. Expert policy groups cannot play this role without risking their insider access. TMP can say things nobody else will say.

• Provoking revealing responses. The OpenAI subpoena campaign was itself the most damning evidence against OpenAI's governance -- that it treats all critics as potential conspirators. TMP's existence forced OpenAI to show its hand.

Weaknesses and Risks

• No demonstrated behavioral impact. After two years, no company has demonstrably changed behavior due to TMP pressure. Johnston acknowledges a "weak track record." The theory of change depends on producing change, and there is no documented instance of it.

• Fundamental sustainability crisis. The adversarial model creates a funding paradox: the more effective TMP is, the more powerful enemies it makes, the harder it becomes to fund and insure. This is not a problem animal welfare bad cops face at the same scale. If TMP cannot solve this, it will not survive long enough to matter.

• Single point of failure. One FTE, one leader, one source of institutional knowledge. If Johnston burns out, gets priced out, or faces legal consequences, TMP effectively ceases to exist.

• The bounded cost problem remains unresolved. Johnston identified this in 2023 and has not articulated a solution. If AI compliance costs are truly unbounded, the entire corporate campaigns theory of change fails regardless of how much pressure TMP generates. The pivot to legal/regulatory action may be an implicit acknowledgment that the original theory was wrong.

• Own transparency gap. TMP demands governance transparency from OpenAI while providing limited transparency about its own board, team, financials, and decision-making. An EA Forum commenter flagged this directly on TMP's launch post.

Cross-References

• Encode (AI policy nonprofit, 3 FTEs): Parallel target of OpenAI subpoenas. Worked on SB 53. Co-organized open letter with TMP. Similar David vs. Goliath dynamic but more focused on policy drafting than adversarial campaigns.
• PauseAI: Johnston sits on their board. PauseAI is more radical in asks (training moratoriums) but less confrontational in tactics than THL was. TMP is the institutional pressure layer that PauseAI's protest model lacks.
• ControlAI: Johnston explicitly critiqued their approach on EA Forum -- concerned they incentivize companies to stop doing good-but-inadequate things. TMP tries to avoid this by focusing on broken promises rather than inadequate-but-existing commitments.
• Tech Oversight Project: Co-authored OpenAI Files. Funded by Omidyar/Hughes ($782K budget). Provides a partner with resources TMP lacks. Relationship depth unclear.
• The Humane League: TMP's spiritual ancestor. Johnston directly imported THL's corporate campaigns model. Key difference: food company compliance costs are bounded; AI company compliance costs may not be.

What Would Change This Assessment

• A documented case of company behavior change attributable to TMP pressure would be the single most important update. If OpenAI, Anthropic, or any target demonstrably changed a policy, practice, or disclosure in response to TMP's work, that would validate the theory of change.
• Sustained funding at $200K+/year would signal that the sustainability problem is solvable and TMP can build institutional durability.
• Successful regulatory action from TMP's IRS complaint or SB 53 allegation would validate the legal/regulatory pivot.
• Johnston articulating a resolution to the bounded/unbounded cost problem would demonstrate strategic evolution.
• TMP ceasing to exist within 2 years would confirm the sustainability critique.

Self-Critique

• Weakest claim: My suggestion that TMP's pivot to legal/regulatory action represents an implicit acknowledgment that corporate campaigns failed. It could equally be natural evolution or strategic diversification, not retreat.
• Potential bias: I may be too sympathetic to TMP because the subpoena story is compelling and OpenAI is an easy villain. The lack of demonstrated impact is a serious concern that should weigh more heavily.
• What I missed: The FLI podcast transcript (unavailable) could contain Johnston's most developed thinking on strategy. The Fast Company profile (inaccessible) may contain important details about TMP's approach.
• Information that would most change my view: Evidence that the SB 53 allegation led to a formal investigation, or evidence that any company changed behavior in response to TMP's work.
• What a thoughtful disagreer would say: "TMP produces accountability journalism, not accountability. Reports and complaints do not reduce AI risk. The theory of change has a missing step between 'publish investigation' and 'company changes behavior.' Animal welfare campaigns worked because they had a proven escalation path (demonstrations, boycotts, regulatory pressure) that TMP cannot replicate against trillion-dollar AI companies."