Below, Hamilton Mann shares five key insights from his new book, Trias Algorithmica: What Code Rules.
Hamilton is an AI researcher, Chairman, and President of the Artificial Integrity Institute, a nonprofit organization dedicated to advancing AI systems that prioritize integrity over intelligence.
What’s the Big Idea?
AI introduces a sixth power: algorithmic power. As it increasingly shapes what we see, think, and value, the question is no longer simply how intelligent AI can become, but what makes its growing power legitimate—and how we prevent it from becoming concentrated and unchecked.
Listen to the audio version of this Book Bite—read by Hamilton himself—in the Next Big Idea App, or buy the book.
1. Mind the sixth power.
We’ve learned to balance and stabilize five power dynamics:
- The Legislative Branch
- The Executive Branch
- The Judicial Branch
- The Media
- Information Technology
Information technology—the fifth power—emerged with the internet and evolved through social networks, democratizing influence by turning Andy Warhol’s “fifteen minutes of fame” prediction into a reality.
AI introduces a sixth power: algorithmic power. What you see is your child asking a chatbot questions, a hiring system ranking your candidate, a bank flagging your transaction, a hospital assigning you a level of risk, a social platform deciding which of your words travel and which don’t, a generative model drafting you a memo, a love letter, a legal argument. All of this looks like technological performance. None of it looks like what is really being produced and accumulated: not intelligence, but power.
Society has learned over centuries how to stabilize the five preceding power dynamics. The greatest challenge AI poses is understanding and governing the new constitutional category of power its presence introduces into society.
2. What is algorithmic power?
Algorithmic power is trained on a specific corpus and objective function. So it ends up encoding values, limits, and permissions as law in the guise of code. Which makes it a kind of legislative power.
It exercises discretionary authority over the allocation of attention, the filtering of reality, the classification of people, the recommendation of actions, the constraint of choices, the pricing of risk, and the distribution of opportunities. Which makes it a kind of executive power.
It sets the standards, applies them, and then judges both itself and everyone subject to those rules. Which gives it a kind of judicial power.
It governs our lives, reconstructing the very concentration of powers that constitutional thought was designed to prevent. It mediates how billions of people perceive the world, their place within it, and what counts as true or false, also subjecting institutions to the terms of an algorithmic order. It exploits human time, interests, thoughts, and creations without any compensation to the point that the term “user” truly defines what it is, rather than the person who thinks they are using it.
“No single institution should write the rules, enforce them, and judge itself.”
As if that were not enough, it does not stop there. It does something even more disturbing: it convinces us that it has commoditized knowledge and expertise. So we are searching for a “human competitive advantage” against it. Looking for qualities that we humans still believe it cannot easily replicate, whether it is about judgment, curiosity, creativity, critical thinking, or leadership. But how can we remain capable of freely exercising these qualities when trapped in the belief that it has commoditized knowledge and expertise?
Because of this, its power also lies in shaping the very conditions under which anything else becomes thinkable in the first place. And we keep asking it for more. In fact, we race for it. Which makes it all the harder to ask: What makes algorithmic power legitimate?
We would never accept legislative, executive, and judicial powers concentrated in government, but as things stand, we accept that concentration in AI. No single institution should write the rules, enforce them, and judge itself. This principle established the foundation of the separation of powers among the three state-level powers, under the intellectual heritage of Montesquieu’s Trias Politica.
The separation of powers has been one of the most enduring innovations in human societies. A true anti-tyranny technology designed to prevent the harmful consequences that arise when these powers accumulate and concentrate in the same hands. Algorithmic power deserves the same constitutional scrutiny that political power has faced for centuries.
3. The presumption of legitimacy.
Unlike traditional forms of domination, algorithmic concentration of power imposes itself through utility, not through force. It arrives as a service. That’s why we may forget to question its legitimacy.
This is the defining characteristic of software absolutism, which compounds through the self-reinforcement of two factors, particularly present in frontier AI systems: convenience through optimization and conformity to patterns.
Convenience through optimization creates dependencies on systems that are incredibly useful but increasingly difficult to question. It speaks the language of personalization, user-friendliness, efficiency, and productivity.
It does not ban alternatives so much as bury them beneath friction. It does not always punish dissent. Sometimes it simply optimizes dissent out of visibility. But when algorithms shape what we see, think, and trust, freedom can erode without anyone noticing. This is “the dependautonomy effect”: the paradox in which technology gives us the feeling of greater autonomy while simultaneously increasing our dependence on the systems that shape our livelihoods. So we come to believe that we are becoming more empowered, while progressively losing the capacity to function independently of the very technologies that empower us.
“When algorithms shape what we see, think, and trust, freedom can erode without anyone noticing.”
Conformity to patterns is another, equally concerning factor contributing to the rise of software absolutism. AI systems are fundamentally pattern-recognition machines. Yet many of humanity’s greatest breakthroughs—scientific discoveries, artistic revolutions, entrepreneurial innovations, social progress—have emerged precisely from the ability to think beyond prevailing patterns, challenge conventions, and imagine alternatives that did not yet exist. As AI becomes more influential in everyday life, societies may become increasingly engineered for conformity to past behaviors rather than originality. And the more we delegate judgment to systems trained on what already exists, the harder it may become to recognize what does not fit established patterns, making new options harder to imagine and the difference harder to value.
4. The F-Curve.
More and more voices are calling for ways to counter the expansion of algorithmic power: asking for pauses so that the economic system can adapt, asking for incentives so that it could be beneficial for all and not further undermine equality, asking for institutional guardrails so that it could be safe, asking for protected human domains to avoid job destruction at scale.
All of these manifestations connect to the cycle we are currently navigating, which I call the F-curve.
Most people think AI will change work by replacing jobs, but I believe that AI is changing the very definition of what society considers valuable work. When AI systems consistently deliver results that are “good enough” at unprecedented speed and cost, they become the benchmark against which human performance is measured. Over time, people do not simply compete with machines; they adapt themselves to standards established by machines. What begins as a tool becomes a reference point for employability, productivity, pricing, and professional legitimacy.
“AI is changing the very definition of what society considers valuable work.”
The resulting pattern resembles the letter F: an initial shift, followed either by stagnation and eventual dissipation along the lower branch of the F, or by a sharp structural acceleration that culminates in a long phase of consolidation along the upper branch. We are currently climbing that upper branch.
Beyond the risk of human work becoming obsolete, the real threat is that algorithmic power, concentrated in the hands of a few, increasingly defines human worth according to benchmarks that we did not collectively create, that no democratic institution has meaningfully scrutinized or anticipated, and that we may collectively lose the ability to contest because they evolve faster than our institutions can adapt. This is one of the reasons why addressing what makes algorithmic power legitimate is essential.
5. Algorithmic power separation.
Establishing the legitimacy of algorithmic power requires determining how that power ought to be divided and restrained so that no algorithmic power can become the judge of its own rules, executor of its own decisions, and final arbiter of its own legitimacy. For that, eight algorithmic power unbundlings apply.
At the level of the algorithmic legislative-like power:
- Goal-setting must be separated from machine-making, so that those who build the system do not also define its aims, success criteria, and the harms it tolerates.
- Data stewardship must be separated from developer control, so that those who build the system do not decide unilaterally what data is collected, how it is used, what is excluded, and who has recourse.
- Training governance must be separated from optimization, so that optimization objectives do not quietly become policy without independent oversight.
At the level of the algorithmic executive-like power:
- Compute governance must be separated from the ability to scale at will, so that access to infrastructure does not become unilateral power over what can be built and deployed.
- Deployment authorization must be separated from design ownership, so that the same team cannot both build a system and decide when, where, and at what scale it enters society.
At the level of the algorithmic judicial-like power:
- Societal impact monitoring must be separated from engagement optimization, so that those optimizing the system are not also the ones defining what success looks like.
- Contestability must be separated from algorithmic finality, so that an algorithmic decision does not become final simply because it is fast, scalable, convenient, and assumed to be neutral.
- Market portability standards must be separated from vertical lock-in, so control over infrastructure cannot become control over everyone else’s models, data, deployment paths, and access to customers.
These eight algorithmic unbundlings undo the primary couplings through which algorithmic power becomes concentrated and self-contained.
Unbundling is necessary for separation, but it is not enough for the creation of meaningful counterpower.
Enjoy our full library of Book Bites—read by the authors!—in the Next Big Idea App:
