Gates is wrong to think that the answer to AI’s growing power is to build institutions capable of governing it. Any institution entrusted with governing power becomes, by design, a concentration of power itself. The answer to concentrated algorithmic power cannot therefore be another concentration of power left constitutionally undivided. How, then, do we prevent the power created to govern algorithmic power from reproducing the very concentration it is meant to constrain? The constitutional question must come first:

What makes algorithmic power—and the power created to govern it—legitimate?

That distinction matters. The challenge before us is to determine how algorithmic power itself should be structured. Gates is right to argue that existing institutions were not designed for a technology that cuts simultaneously across employment, education, security, taxation, health, infrastructure, and governance. But creating new institutions, even where necessary, does not answer that prior question of legitimacy. It only makes the need for a constitutional architecture of algorithmic power more urgent.

Algorithmic power should no longer be understood primarily as a question of intelligence. It has become a question of governance power. Algorithmic systems increasingly determine what is visible, what is recommended, who is ranked, who is trusted, who is denied, and which choices appear possible in the first place. They don’t just deliver services. They increasingly mediate the conditions under which we, as individuals, corporations, and institutions, both public and private, perceive and act upon the world. Once systems acquire that capacity, the relevant question is no longer only whether they are safe, accurate, or useful. The prior question is one of legitimacy. Only from that question follow the others: where that power is located, who exercises it, under what constraints, and who has the authority to challenge it.

The central challenge is that contemporary advanced algorithmic systems fuse powers that societies learned long ago to separate. The same organization can define the objectives of a system, determine the data on which it learns, build and deploy it at scale, establish the benchmarks by which it will be evaluated, and then judge whether its own system is safe enough to remain in operation. Rulemaking, execution and judgment collapse into a single institutional chain. An arrangement that would trouble us deeply in almost any other political setting has become routine in technology.

So the answer to Gates’s call for governance cannot be more governance. It must be governance redesigned against the concentration of governance itself.

That is the constitutional intuition behind Trias Algorithmica .

Algorithmic power must not be allowed to govern itself. No actor should possess unilateral control over the rules encoded into a system, the operational capacity to deploy those rules at scale, and the authority to determine whether their consequences are acceptable. This requires something more demanding than principles, voluntary commitments or corporate ethics boards. It requires independent counterpowers capable not merely of advising, but of contesting, delaying, reviewing and, where necessary, overruling algorithmic authority.

Constitutionalism Is A Counterpower

Seen through that lens, several of Gates’s proposals become even more consequential.

His idea of a Human Reserved domain is a question of legitimate boundary-setting. Who has the authority to decide that a particular human function should not be automated, even when automation is technically possible and economically efficient? Of course, that decision cannot legitimately, or economically, belong to any actor acting in isolation.

A company that decides alone to preserve a function for humans may simply impose a cost disadvantage on itself while competitors automate the same activity, turning an ethical commitment into a competitive penalty. An organization that deploys an AI system cannot set such boundaries independently either, because its choices are embedded in supply chains, labor markets, procurement ecosystems, and competitive pressures that extend far beyond its own perimeter. A city or state acting alone may protect certain forms of human work locally, yet simultaneously encourage firms, investment, or automated activities to relocate to neighboring jurisdictions where those constraints do not apply. And even a country acting unilaterally faces the same problem at a larger scale: divergent rules can produce regulatory arbitrage, offshoring, distorted trade, uneven production costs, and ultimately a race toward the jurisdiction imposing the fewest constraints on automation.

The economic problem is therefore inseparable from the constitutional one. A right that can be economically bypassed simply by moving the activity, the capital, the compute, or the deployment elsewhere is not yet an effective right. If certain domains of human activity are deemed important enough to remain Human Reserved, their protection cannot depend solely on the willingness of one company, one institution, one city, one state, or even one country to bear the economic cost of preserving them while others do not. Such boundaries require coordination at a level commensurate with the markets, technologies, and capital flows they seek to govern.

They therefore rise to the level of constitutional rights and, in some cases, potentially universal rights. This is precisely the kind of problem Trias Algorithmica addresses through the separation of goal-setting from machine-making: those who possess the technical capability to create or deploy a system should not automatically possess the authority to determine its purposes, acceptable harms, or the social boundaries within which it operates. But the same principle must also extend across jurisdictions: legitimate limits on algorithmic power must be designed so that no actor can unilaterally impose them, evade them, or gain a structural economic advantage simply by refusing to recognize them.

The same is true of deployment. Gates is right to insist that society needs time to adapt, but the idea that such time might be created simply by slowing AI development is ultimately too optimistic. Indeed, Gates himself acknowledges the weakness of that proposition: the geopolitical, competitive, and economic forces driving AI forward make coordinated global restraint extraordinarily difficult to sustain. Any company that slows while its competitors accelerate risks losing market position; any country that imposes unilateral limits risks shifting investment, talent, compute, and strategic advantage elsewhere.

Undivided Power Cannot Be Slowed

A voluntary slowdown therefore depends on precisely the kind of collective self-restraint that concentrated power has historically proved least capable of sustaining. The answer cannot be to hope that algorithmic power will voluntarily advance more slowly. Power does not reliably restrain itself; left unchecked, it tends to seek continuity through expansion until it encounters a counterpower. The task, therefore, is to ensure that, however quickly technological capability advances, no actor possesses the unilateral authority to convert that capability into concentrated algorithmic power capable of governing society.

Trias Algorithmica gives this principle an institutional form: deployment authorization for consequential algorithmic systems should be separated from design ownership. The organization that creates a powerful algorithmic system should not alone determine when it enters society, at what scale, in which domains, and under what conditions. The objective is not to slow innovation for its own sake, but to subject the conversion of technological capability into societal power to independent authorization, contestability, and review. Rather than asking technological development to wait for society, we should build a governance architecture capable of governing technological development at speed. That is the deeper purpose of algorithmic separation of powers: not to rely on the self-restraint of those who hold power, but to make restraint structural by confronting power with counterpower.

And this is also where governance must become capable of learning without surrendering authority. Decisions about AI will inevitably be made under uncertainty. Waiting for perfect evidence is not neutrality; it simply allows existing technological and market forces to determine the outcome by default. But acting under uncertainty does not require pretending to possess certainty. Governance can be designed to remain revisable.

Trias Algorithmica proposes precisely such a mechanism through independent societal signal monitoring. Authorization should not be understood as a one-time verdict. Systems that have been allowed into society should remain subject to continuous scrutiny, just as medicines remain subject to pharmacovigilance after approval. Unexpected harms, whether discrimination, psychological dependency, labor disruption, security vulnerabilities or broader social effects, should be capable of triggering investigation, correction, restriction or suspension. The important principle is that learning does not remain under the sole control of the organization being evaluated.

This gives us a different way to think about uncertainty. Legitimacy does not require that governance always be right. It requires that governance be contestable, revisable and distributed enough that error does not become irreversible authority.

That point may ultimately be more important than any individual regulatory proposal.

The temptation in responding to AI’s growing power is to assume that creating institutions capable of governing it is enough to make that governance legitimate. But institutional capacity and institutional legitimacy are not the same thing. The category error is not that Gates calls for institutions that are too weak or too powerful; it is to assume that institutional capacity itself confers legitimacy. It does not. Legitimacy depends on how power is divided, constrained, contested, and overruled. That is the constitutional problem Trias Algorithmica is designed to address.

This is why the institutions Gates calls for are necessary, but not sufficient.

The Constitutional Promise of Trias Algorithmica

The real task is to ensure that whatever institutions we build reproduce neither technological absolutism nor regulatory absolutism. The goal cannot be to transfer concentrated power from Silicon Valley to governments and declare the constitutional problem solved. The goal must be to make algorithmic power divisible: to separate the authority to define purposes, control data, optimize systems, authorize deployment, monitor consequences, adjudicate disputes, govern compute and prevent infrastructural lock-in.

Gates is therefore right that the world needs a plan. But the deeper question is what kind of plan can remain legitimate as both the technology and our knowledge of its consequences evolve.

The answer may lie less in predicting the future correctly than in designing an architecture in which no actor has the power to make its own prediction final.

That is the constitutional promise of Trias Algorithmica. Not to prevent algorithmic power from existing, but to prevent it from becoming absolute; not to diminish the usefulness of algorithmic systems, but to make their power contestable; and not merely to govern what AI can do, but to govern who gets to decide what it should do, and on what basis that power is legitimate.

History taught us that the concentration of rule-making, executive, and adjudicative powers required Trias Politica: a separation of powers. Algorithmic systems have reconstructed that same concentration of power within the stack. The algorithmic age we are now living through requires its own separation of powers—the separation of algorithmic power. Trias Algorithmica: the constitutional answer to what code rules.

Trias Algorithmica is a constitutional framework in the lineage of Montesquieu’s Trias Politica , designed to separate algorithmic power so that no single actor can define, execute, and self-judge the algorithmic rules shaping society. More in Trias Algorihtmica: What Code Rules.