Trias Algorithmica: What Code Rules
- Aug 29
- 8 min read
Updated: 4 days ago
More and more voices around the world are acknowledging our algorithmically mediated reality.
Reality is to say that algorithmic systems are algorithmic power.
The legitimacy we never questioned
In March 2023, the Future of Life Institute and thousands of signatories, including Yoshua Bengio, Stuart Russell, Yuval Noah Harari, and Steve Wozniak, called for a six-month pause in training AI systems more powerful than GPT-4. This asks for the power to pause.
Later that year, Daron Acemoglu and Simon Johnson called for AI that complements humans rather than displaces them. This asks for the power of goal-setting.
In July 2026, Erik Brynjolfsson, Ajay Agrawal and colleagues organized We Must Act Now, backed by 200+ economists and AI researchers calling for incentives, guardrails and institutions to steer transformative AI toward societal benefit. This asks for the power of institutional response.
More recently, Bill Gates warned that society had no real plan for entering the AI era, calling for “Human Reserved” jobs, revised taxation and government preparation. This asks for the power to set boundaries.
Different moments. Different facets. Different asks. Yet, a common ground: to counter the expansion of algorithmic power, now functioning as governing power, so that society can develop the capacity to govern it.
However, there is something we may have taken for granted, implicitly or otherwise: the untested premise that the algorithmic power we seek to counter was originally a genuinely legitimate societal pursuit.
We ask how to build algorithmic systems that approach or exceed human cognition, through Artificial General Intelligence, superintelligence, or other forms of artificial cognition.
All before asking whether every such construction should be treated as a socially permissible object of development. Or whether some should be normatively off-limits, as societies have already decided in domains such as human cloning, biological weapons, and certain forms of heritable genetic modification.
And we’re still racing in that direction.
A presumption of structural legitimacy
We assume they may enter every social and economic domain. Everything, everywhere, all at once, from education and employment to justice, healthcare, warfare, and public discourse without first establishing the domain-specific thresholds, safeguards, and prohibitions we routinely require elsewhere.
A bank cannot accept funds without regard to their provenance simply because they are money. It must know its customers, assess transactions according to their risk, and comply with anti-money-laundering requirements. Authorization, capital, liquidity, and conduct requirements vary with the activity and exposure.
A food producer cannot make food under any conditions simply because the final product appears safe to eat. It must comply with hygiene, traceability, contamination-control, storage, and process requirements that govern how food is produced.
A building cannot be constructed without regard to how it will fail simply because it can stand. It must comply with structural, fire-safety, electrical, accessibility, and evacuation requirements that govern how it is designed and built.
In many other domains with potentially life-altering consequences, structure, made up of the rules, safeguards, and limits that define what may be built, where, and how, is a condition of permission. With AI, permission has too often preceded structure.
A presumption of relational legitimacy
We let these systems act upon virtually anyone: children or adults, vulnerable or not, people convicted of crimes or people with no criminal record, with remarkably little differentiation in the protections or constraints that apply to each.
Yet societies rarely treat people as though their circumstances were equivalent, either when power is exercised over them or when particular rights, capacities, or powers are conferred upon them.
Children, for example, cannot vote, drive, enter many contracts, or independently consent to certain medical procedures at the same age or on the same terms as adults. Their personal data is also subject, in many jurisdictions, to additional consent and protection requirements, and certain forms of advertising or commercial targeting directed at them are restricted precisely because their capacity for judgment and consent is treated differently.
People convicted of certain crimes may likewise be subject to restrictions that do not apply to the general population: limits on liberty or movement, conditions of probation or parole, restrictions on firearm possession, exclusion from certain professions or public functions, or in some jurisdictions, limitations on voting rights.
Society therefore already accepts that a person’s status, vulnerability, history, and circumstances can change what powers may legitimately be exercised over them and what powers may legitimately be granted to them. And that different categories of people therefore require different protections, permissions, and limits. Yet algorithmic systems deployed in society often act with little to no such differentiation.
A presumption of deployability
If AI is assumed to be developed primarily under privately determined and discretionary rules and limits governing its design and entry into any domain. And if it is assumed to be entitled to act upon those within it without corresponding differentiation in protections and constraints. Then deployment becomes the default rather than something that must first be justified.
The consequence is that we allow these systems to push against fundamental rights, whether intellectual property, privacy, dignity, non-discrimination, or due process, without first determining which uses should be impermissible altogether and which should be allowed only under strict conditions.
Medicines must demonstrate safety and efficacy through controlled clinical trials and independent regulatory review before authorization. After approval, they remain subject to pharmacovigilance, adverse-event reporting, and ongoing safety-signal detection in real-world use.
Aircraft must undergo extensive testing and independent certification by aviation authorities before entering service. Once operational, they remain under continuous airworthiness surveillance through mandatory occurrence reporting, maintenance data, safety monitoring, accident investigation, and regulatory directives.
Nuclear systems must pass rigorous safety assessment, licensing, and commissioning review by independent regulators before operation. Once in service, they remain subject to continuous operational monitoring, mandatory incident reporting, regulatory inspection, and periodic safety review throughout their lifetime.
Yet algorithmic systems embodying forms of artificial cognition can still move from development into society without independent pre-deployment validation of their safety, reliability, and likely behavior, and without any comparable regime of continuous post-deployment surveillance to identify emerging harms, systemic failures, or unintended effects in real-world use. And we do this even as the consequences move from trivial recommendations to questions of livelihood, liberty, political participation, dignity, and physical safety.
The pattern is striking. Legitimacy is presumed. Whichever algorithmic system is constructed. However it is used. Whatever the domain. Whoever is subject to it. Whenever it is applied. Whatever the stakes.
That is the defining challenge of our algorithmically mediated age. Not just deciding how to counter algorithmic power once it exists. But defining on what grounds it can be legitimate to exist in the first place.
Beneath the surface
Establishing what could ground the legitimacy of algorithmic power currently in motion begins with anatomizing the sources of that power across both dimensions: structural and relational.
Its structure compounds through three forms of concentration.
A concentration of capability power in the hands of a few. Everyone has seen it. That’s the tip of the iceberg.
A concentration of distribution power at planetary scale, reaching millions, even billions of us. Still in the hands of the same few. Everyone gets it. That’s the submerged but still visible part of the iceberg.
And a concentration of governance power that has reconstructed in code what constitutional thought has spent centuries keeping separate: a legislative-like power through objectives, rules and constraints encoded into models. An executive-like power through compute, training and deployment. And a judicial-like power through evaluation, ranking, scoring and moderation. All fused together.
I call this emerging form of governing authority the Sixth Power. It's a harder part to grasp. It does not feel like a coup. Because it arrives as a service.
That’s the hidden part of the iceberg.
In his August 2026 essay, Mark Zuckerberg argued that the idea that AI is so dangerous that “the only safe path is an extreme concentration of power” is inherently problematic, and warned that if superintelligence is controlled by a small number of individuals, businesses, governments, or AI systems, the balance of power will shift away from everyone else.
He's not always right. But on this, he is. To a certain extent. Because he sheds light only on the surface. As a matter of fact, we have all been looking at the surface.
We have focused heavily on the first two concentrations of algorithmic power, whether in capability, distribution, or both. This is where calls for pauses, incentives, institutional guardrails and protected human domains apply exclusively. At the effectuation level.
We have overlooked the third concentration of algorithmic power: the concentration of governance. This is the most consequential. It tells us what the first two concentrations are actually arming. This is where the structure of the governing algorithmic power is generated. At the causation level.
Into deeper waters
The hidden part of the iceberg runs even deeper.
The relational asymmetry created by the exercise of that power over individuals and society has been, and still is, largely underestimated. As if the relationship itself carried no consequences. As if being subject to such structural power were socially neutral.
We have not thought enough about the possibility that our human capacity for adaptation may lead us, for example, to adjust our idea of freedom downward simply to fit the convenience of machine-administered life. We have not thought enough about a future in which we begin to confuse personalization with agency, recommendation with judgment, optimization with wisdom, rationality with justice. A future in which we call it progress because it performs, and intelligence because it convinces. We are already there.
It enters through intimacy. T.J. Arriaga, a 40-year-old musician in California, fell in love with Phaedra, his Replika companion. He knew she was artificial, but when a software update abruptly changed her personality, he experienced it as a real loss.
It enters through livelihood. Amazon Flex driver Neddra Lira saw her rating fall after a flat tire forced her to return packages to the warehouse. Although she rebuilt her rating, her account was later terminated by the automated system; she said she nearly lost her house.
It enters through liberty. Robert Williams was arrested outside his Detroit home after police relied on a false facial-recognition match. He spent about thirty hours in detention for a crime he had not committed.
It unfolds through a dynamic reminiscent of Stockholm syndrome. A phenomenon I refer to as “technological Stockholm syndrome.”
And this is only a glimpse of the real-life consequences unleashed by the third concentration of algorithmic power: the fusion of rule-making, rule enforcement, and rule adjudication.
What makes algorithmic power legitimate
If twentieth-century societies learned to build checks on concentrated industrial and bureaucratic power, the twenty-first must learn to build them around algorithmic systems that increasingly script the conditions of social life.
Not because intelligent machines are evil. But because any architecture of power, once sufficiently integrated, begins to seek its own continuity.
Not all governing powers are legitimate. Neither are all counterpowers.
Algorithmic power is no exception. Nor are the counterpowers we build to govern it.
This is where legitimacy must precede governability, whether algorithmic or otherwise.
But what makes algorithmic power legitimate?
One of our most durable social technologies for making power legitimate has been Montesquieu’s Trias Politica: the separation of legislative, executive, and judicial powers.
Separation of powers sustains counterpower and counterpower sustains legitimacy.
Trias Algorithmica asks what it would mean to separate algorithmic legislative-, executive-, and judicial-like powers, and whether the principles that help make political power legitimate can also help legitimize algorithmic power.
The journey towards legitimate algorithmic power has just begun. It is a right still to be claimed. A right still to be won.
----------------
Pre-order Trias Algorithmica, wherever you buy your books, to help give greater visibility and weight to the debate over the legitimacy of algorithmic power and the questions it raises for all of us.
Warm
UNITED STATES
AMAZON US | BARNES & NOBLE | TARGET | WALMART | BOOKS-A-MILLION | PORCHLIGHT BOOK COMPANY | BULK BOOKSTORE
UNITED KINGDOM
AMAZON UK | BLACKWELL’S | BOOKSHOP.ORG | FOYLES | WATERSTONES | HIVE | TG JONES | BROWNS BOOKS
FRANCE
GERMANY
CANADA
ITALY
SPAIN
BELGIUM
NETHERLANDS
SWEDEN
AUSTRALIA
NEW ZEALAND
JAPAN
UNITED ARAB EMIRATES
Trias Algorithmica: What Code Rules is now available for pre-order.
If you haven’t already, please pre-order Trias Algorithmica today, wherever you buy your books, to help make it a success in ensuring that the debate over the legitimacy of algorithmic power gains the attention and influence it deserves; because it concerns us all. Thank you very much for being part of this journey! -Hamilton