How Authority Consolidates
The Pattern of Consolidation
OpenAI launched in 2015 as a nonprofit with a clear mission: develop artificial intelligence for the benefit of all humanity. When capital demands forced the creation of a for-profit subsidiary in 2019, the nonprofit board retained full control. Investors could earn returns, but those returns were capped, and the nonprofit’s mission took precedence. It was an unusual arrangement, designed to distribute authority deliberately so that the people funding the work would not control its direction. In December 2024, the company announced it would restructure that governance.
The December announcement proposed undoing the structure. The nonprofit would step aside, and the for-profit arm would convert to a public benefit corporation accountable primarily to shareholders. The justification was straightforward: OpenAI needed billions more for infrastructure, and the nonprofit structure was limiting access to capital.
The backlash came quickly. By April 2025, an open letter from former employees, academics, and civic leaders warned that the restructuring would violate OpenAI’s founding commitments. State attorneys general in California and Delaware launched investigations. In May, facing mounting pressure, OpenAI reversed course and announced that the nonprofit would retain control after all.
When the restructuring closed in October 2025, the numbers told a different story. The nonprofit held 26 percent equity while Microsoft held 27 percent. The nonprofit still formally appoints the board, but the operational authority had shifted. The caps on investor returns were gone. What began as distributed governance, with mission explicitly prioritized over profit, had consolidated around capital. The structure designed to prevent exactly that outcome could not sustain itself under pressure.
This was not a story of deception or bad actors. No one lied about their intentions. The problem was structural, not personal. Systems designed to distribute authority tend toward consolidation anyway, not through conspiracy but through the mechanics of how power moves when resources are constrained and stakes are high.
The pattern is not unique to OpenAI. Six months earlier, in June 2025, the Trump administration rebranded the US AI Safety Institute. The institute had launched in November 2023 under Biden with a clear mandate: evaluate AI systems for risks, develop safety standards, facilitate international cooperation on responsible AI development. The framing emphasized diverse stakeholder input, with over 200 organizations in its consortium, and a mission focused on making AI “safe, secure, and trustworthy.”
The rebrand changed the name to the Center for AI Standards and Innovation. The shift was more than cosmetic. The new mission emphasized guarding against “burdensome and unnecessary regulation of American technologies by foreign governments” and ensuring “US dominance of international AI standards.” The multi-stakeholder consortium remained, but the purpose had moved from safety evaluation to protecting corporate interests. What had been framed as distributed input serving the public good became centralized authority serving national competitive advantage. The same week the Safety Institute rebranding was announced, the federal government launched Operation Metro Surge in Minneapolis, deploying over 2,000 immigration enforcement agents following fraud allegations. When an ICE agent shot Renee Good while she observed the operation, federal authorities blocked state investigation and Congressional oversight. The justification was urgency and public safety. The result was consolidated federal control that overrode distributed checks at state and local levels.
The same mechanics appear across sectors. Participatory grantmaking initiatives launch with community members making funding decisions, then shift to “advisory” roles when foundations cite capacity concerns. Global health governance starts with country coordinating mechanisms that include community representation, then consolidates decision-making to international consultants and donors as technical requirements and reporting burdens increase. Evaluation bodies are chartered with independent oversight authority, then find that authority limited to advisory status when their conclusions threaten organizational priorities.
The question is not whether consolidation will happen. The question is why it happens so reliably, and whether different structures can resist it.
Agency, Plurality, Reciprocity
Nate Silver offers one framework in On the Edge. His book maps a cultural divide between what he calls “the River” and “the Village.” The River is the domain of gamblers, Silicon Valley founders, and others who think probabilistically and take calculated risks. The Village is the East Coast establishment: media, academia, government institutions that prioritize consensus and risk aversion. Silver argues that the River succeeds because its members maximize expected value, think in probabilities, and maintain competitive independence rather than seeking group consensus.
For governance in the AI era, Silver proposes three principles: Agency, Plurality, and Reciprocity. Agency means maintaining the ability for individuals and groups to develop their own options for action rather than having decisions made for them. Plurality means actively preventing any single person, group, or ideology from gaining monopoly control over decision-making. Reciprocity means treating others as intelligent and capable of strategic behavior rather than as subjects to be managed.
These principles apply beyond AI. Plurality is particularly relevant to the consolidation problem. When decision-making authority consolidates, information is lost. A single executive cannot hold what a distributed network knows. A centralized committee cannot see from as many angles as independent actors pursuing their own strategies. Consensus-building smooths over disagreements that might contain crucial information. The pursuit of unified direction eliminates the optionality that comes from maintaining multiple competing approaches.
This contrasts with how the Village operates. Decision-making flows through consensus, which means waiting until everyone agrees or until dissenting voices are pressured into silence. The result is risk aversion that masquerades as wisdom. Decisions get made not because they are best but because they are least objectionable to the most people. The process values harmony over accuracy, group cohesion over individual agency.
OpenAI’s restructuring violated all three. The nonprofit board’s agency to control the mission was stripped when investor pressure overrode governance structure. Plurality disappeared when capital concentration gave Microsoft and a small group of funders effective control over strategic direction. Reciprocity broke down when the restructuring proceeded despite vocal opposition from the civic leaders, state officials, and former employees who had stakes in the outcome. What remained was a structure that concentrated decision-making authority precisely where the original design had tried to prevent it from going.
The AI Safety Institute rebranding followed the same pattern. Agency shifted from safety evaluators to those optimizing for national competitive advantage. Plurality narrowed from multi-stakeholder input to a single framing: protect US corporate interests. Reciprocity was replaced by a presumption that safety concerns were obstacles rather than legitimate considerations requiring serious response.
Why Systems Consolidate
The question is why this happens so reliably. Silver’s framework names what distributed authority requires, but not what pulls against it. The forces favoring consolidation are structural, not personal.
The first is efficiency. Distributed decision-making is slow. Consulting stakeholders takes time. Building consensus requires patience. Getting buy-in from multiple parties delays action. When urgency is real or can be made to seem real, efficiency becomes the argument that justifies centralizing control. In Minneapolis, allegations of fraud justified deploying over 2,000 federal immigration agents. When an ICE agent fatally shot Renee Good on January 7, the response from DHS wasn't investigation or pause but escalation: hundreds more agents. When state investigators tried to examine the shooting, the FBI blocked them. When Congress tried oversight, DHS required seven days notice. The justification was urgency. The result was consolidated federal authority that overrode every distributed check. Speed justified the override.
The second is accountability. Distributed systems make it hard to identify who is responsible when things go wrong. If authority is shared, blame diffuses. Boards, funders, and oversight bodies prefer clear lines of accountability, which means clear hierarchies, which means someone at the top who can be held responsible. The structure that distributes authority also distributes accountability in ways that make external oversight difficult.
The third is resource concentration. Whoever controls the money shapes the decisions. OpenAI’s nonprofit board could appoint directors, but Microsoft held the capital that determined whether the company could build what it intended to build. The formal authority was distributed, but the practical leverage was not. When resources concentrate, decision-making follows.
The fourth is exhaustion. Maintaining distributed authority requires sustained effort. Participatory structures demand ongoing engagement from multiple parties. Consensus-building is labor-intensive. Over time, as people tire or move on, authority consolidates to whoever remains in the room. The path of least resistance runs toward centralization.
What Actually Works
But distributed authority can hold if designed to resist these pressures. Wikipedia offers a model. The platform has maintained genuinely distributed control over knowledge production for over two decades, despite operating at massive scale and facing constant pressure to consolidate.
Wikipedia’s governance distributes veto power rather than just inviting input. No single entity controls what counts as knowledge. Articles are open for anyone to edit, but that openness is structured by policies that no central authority can override unilaterally. Major decisions require community consensus, not executive approval. Different levels of article protection exist for different degrees of contestation, but the decision about which level to apply is itself distributed across editors and administrators who have earned standing through contribution rather than appointment.
The system has formal mechanisms that prevent capture. The Wikimedia Foundation provides infrastructure but does not control content. Corporate sponsors cannot dictate what appears in articles about their companies. Governments cannot remove information they find inconvenient. The barriers to consolidation are technical and social, embedded in how the platform operates rather than depending on any individual’s commitment to maintaining distributed authority.
Wikipedia is messy. Decisions take longer than they would with executive control. Edit wars break out. Some articles sit in contested states for years. The process is inefficient by design, and that inefficiency is the point. Speed would require someone to have final say, which would create the chokepoint that enables capture. Not to mention the funding challenge.
The structure is not utopian. It has problems. Systemic bias exists in who contributes and what gets covered. Power law distributions emerge in who edits most frequently. Informal hierarchies develop among longtime contributors. But the formal distribution of authority has held. No single party can control the platform, and that constraint is structural rather than aspirational.
The principles are portable. Distribute veto power, not just voice. Create multiple paths to the same outcome so that blocking one path does not block all progress. Build costs to consolidation, whether technical, social, or economic, that make centralization harder than maintaining distribution. Make decision-making authority transparent. Rotate power regularly so that no individual or group can accumulate enough standing to override structural constraints.
This is not a governance model that optimizes for speed or simplicity. It optimizes for resilience against capture, which makes it slower and more complex than alternatives. The trade-off is deliberate. Distributed authority survives not by being more efficient but by being harder to consolidate.
Cooperatives offer a different model, one that addresses resource concentration directly. In a cooperative, ownership and governance rights are tied to membership rather than capital investment. The legal structure distributes control by design: one member, one vote, regardless of how much money you contributed. This inverts the usual logic where whoever holds capital shapes decisions.
The model works across contexts and scales. Consumer cooperatives like grocery co-ops and credit unions distribute control among customers. Worker cooperatives give employees ownership and governance authority. Producer cooperatives, particularly agricultural co-ops, allow farmers or other producers to pool resources and negotiate collectively while maintaining distributed control over strategy and operations.
I saw this work in international development contexts where farmers with minimal political or purchasing power on their own could block together through cooperatives to advocate for themselves, negotiate better prices, and improve their economic security. The cooperative structure gave smallholder farmers leverage they could not achieve individually. Decisions about what to grow, when to sell, how to invest collective resources stayed with the members rather than consolidating to whoever provided capital or technical assistance.
But the structure alone is not sufficient. Cooperatives face the same pressures toward consolidation as any other governance model, and they consolidate when the distribution is not maintained through practice. Social hierarchies reassert themselves. Existing power imbalances within communities mean that some voices dominate decision-making even when formal voting rights are equal. Political or ethnic divisions can fracture what is supposed to be collective governance. Gender dynamics shape who speaks in meetings and whose priorities get resourced.
The cooperatives that maintain distributed authority are the ones with clear understanding of cooperative governance principles and the organizational capacity to enact them. This means training members in democratic decision-making, establishing transparent processes for how decisions get made, creating mechanisms for accountability that prevent informal capture by influential members or staff. It requires formalization: written bylaws, regular elections, clear procedures for raising concerns and resolving disputes.
Where capacity is weak or governance mechanisms are poorly understood, cooperatives consolidate despite the legal structure. An elected board becomes a self-perpetuating group. A general manager accumulates decision-making authority that should rest with members. Meetings become performative, with real decisions made elsewhere. The cooperative structure remains on paper while the practice shifts to centralized control.
Land O’Lakes, a large farmer-owned agricultural cooperative, demonstrates that the model can work at scale. The company operates as a major player in the food and agriculture industry while maintaining its cooperative structure. Members retain ownership and voting rights. The board is elected by and from the membership rather than appointed by investors. Strategic decisions flow through democratic processes even as the organization competes in markets where most competitors are investor-owned corporations.
The cooperative model resists consolidation through legal barriers that Wikipedia achieves through technical ones. Converting a cooperative to an investor-owned company requires member approval, and members have strong incentives not to vote away their own control. The structure creates costs to consolidation that make maintaining distribution the path of least resistance.
But like Wikipedia, the cooperative model trades efficiency for resilience. Democratic decision-making is slower than executive control. Member education and engagement require ongoing investment. Consensus-building takes time. The inefficiency is the point. Speed would require centralizing authority, which would undermine the distribution the structure exists to protect.
The principles are the same as Wikipedia but the mechanisms differ. Cooperatives distribute control through ownership structure and voting rights. Wikipedia distributes it through open editing and community consensus. Both create friction against consolidation. Both require ongoing effort to maintain. Both accept slower decision-making as the cost of avoiding capture.
Application to AI Governance
The principles apply beyond Wikipedia and cooperatives. AI governance is the obvious case, given OpenAI’s consolidation and the Safety Institute’s rebranding, but the pattern extends to any domain where distributed authority is claimed but vulnerable to capture.
For AI governance specifically, plurality would mean designing bodies where no single nation, corporation, or ideological camp can dominate. This is difficult because AI development is already concentrated in a small number of wealthy countries and powerful companies. Formal governance structures can invite broad participation while real authority follows capital and compute. Plurality requires distributing not just seats at the table but veto power over decisions that matter. If the AI Safety Institute network wants to resist consolidation toward US or Chinese interests, it needs mechanisms that allow smaller nations or civil society groups to block proposals that serve narrow geopolitical goals. Without that structural distribution of blocking power, “multi-stakeholder” becomes a performance that legitimizes outcomes already determined by whoever holds resources.
Agency in AI governance means preserving the ability of different actors to pursue different approaches rather than converging on a single model or framework. The EU’s AI Act, the US emphasis on innovation and competitiveness, China’s governance model, and approaches emerging from the Global South reflect different values and priorities. A governance structure that maintains agency would allow these differences to coexist and compete rather than forcing harmonization around whichever approach commands the most power. This is uncomfortable for those who want clear global standards, but premature convergence eliminates the optionality that comes from trying multiple paths simultaneously.
Reciprocity means treating other governance participants as capable of strategic behavior rather than as problems to be managed. When powerful nations design frameworks and then invite others to adopt them, that is not reciprocity. When corporations fund governance research and then shape which findings get amplified, that is not reciprocity. When expert bodies position themselves as translating complex technical matters for policymakers presumed unable to understand, that is not reciprocity. When federal agents flood a city, kill a resident, and then prevent state and local authorities from investigating, that is not reciprocity. It is consolidated authority treating distributed governance structures as obstacles to override rather than legitimate participants in accountability.
Genuine reciprocity assumes that people in different positions have different knowledge, that disagreement reflects legitimate differences in values and context, and that governance structures should enable negotiation between equals rather than instruction from center to periphery.
For foundations and development programs, plurality means building governance structures where community advisory boards have binding authority, not just advisory roles. It means creating multiple pathways to funding decisions so that blocking one does not block all. It means transparency about who actually decides and designing friction into processes that attempt to consolidate control.
For any organization claiming distributed governance, the question is not whether the language is present but whether the structure enforces it. Do stakeholders have veto power or just voice? Are there multiple independent paths to achieving goals or does everything flow through a single chokepoint? What are the costs to consolidation? Is decision-making authority transparent enough that informal capture becomes visible? Does power rotate or accumulate?
These are not questions with simple answers. Maintaining distributed authority in practice is harder than proclaiming it in mission statements. The pressures toward consolidation are real, and they do not disappear because a governance document says they should. But the difference between systems that consolidate quickly and those that resist consolidation longer comes down to whether distribution is structural or aspirational.
The lesson is not that distributed authority is impossible but that it requires design. It does not happen by accident, and it does not persist through good intentions. It persists through mechanisms that make consolidation harder than maintaining distribution, through costs and barriers and multiple paths and rotating power and transparency that exposes attempts at capture.
Silver’s framework names what is needed: agency, plurality, reciprocity. The implementation question is whether those principles get embedded in structures that enforce them or remain aspirations that disappear when tested.
The Test
OpenAI began with a structure designed to prevent exactly what happened. The nonprofit board held control precisely so that mission could override profit, so that the people funding the work could not dictate its direction. The structure distributed authority deliberately. It was not enough.
This is not because the people involved were malicious or because the safeguards were poorly designed. It is because the forces pushing toward consolidation are structural. Capital concentrates. Urgency mounts. Efficiency becomes the argument that justifies centralizing control. Exhaustion sets in. The path of least resistance runs toward whoever can act fastest with the most resources.
Distributed authority does not persist through good intentions. It persists through design that makes consolidation harder than maintenance. Wikipedia has held for two decades not because its contributors are more committed than OpenAI’s nonprofit board but because the distribution is technical and social, embedded in how the platform operates rather than depending on anyone’s virtue. Cooperatives resist because legal structures make voting away control difficult and costly.
The AI governance bodies forming now face the same pressures OpenAI faced. The International Network of AI Safety Institutes, national regulatory agencies, public-private partnerships, all launch with language about multi-stakeholder engagement and diverse perspectives. The question is whether that diversity is structural or performative. Do participants have veto power or just voice? Are there multiple independent paths to decisions or does everything flow through a single chokepoint? What are the costs to consolidation? Is authority transparent enough that informal capture becomes visible?
Silver’s framework offers a test. Agency: Can different actors pursue different approaches, or does the system force convergence? Plurality: Does any single nation, corporation, or ideology have monopoly control over decisions that matter? Reciprocity: Are participants treated as equals capable of strategic behavior, or as subjects to be managed?
Most governance structures fail these tests within months of launch. Authority consolidates to whoever holds capital, compute, or geopolitical leverage. The language of distribution remains while the practice shifts. What begins as genuine attempts at shared governance becomes mechanisms that legitimize outcomes already determined by power.
This is not inevitable, but resisting it requires more than aspiration. It requires friction built into the system. Multiple paths that cannot all be blocked simultaneously. Rotation of authority that prevents accumulation. Transparency that exposes attempts at capture. Costs to consolidation that make maintaining distribution the easier path.
The alternative is what we keep seeing: systems designed for plurality that consolidate anyway, not through conspiracy but through the ordinary mechanics of how power moves when resources are scarce and stakes are high. OpenAI is the pattern. The AI Safety Institute’s rebranding is the pattern. The question is whether the governance structures being built now will break it or repeat it.
Anthralytic is a social impact strategy and evaluation studio combines human expertise, data, and AI to help mission-driven teams clarify and amplify their impact.
