The dominant AI risk is not rogue superintelligence but the structural removal of human leverage. As society and warfare run with fewer humans, the elites who once needed 'the many' no longer do — and the contestability that has historically checked power collapses.
The dominant AI risk is not rogue superintelligence but the structural removal of human leverage. As society and warfare run with fewer humans, the elites who once needed 'the many' no longer do — and the contestability that has historically checked power collapses.
The diagnostic isn't 'is the AI aligned' — it's can power still be challenged. Toggle the eight checks. Watch the needle swing between managed populations and distributed agency.
Every room you enter where AI is discussed runs the same loop. What if it becomes conscious? What if it deceives us? What if it decides we are the problem? Those questions are not foolish, and they are not the dominant risk vector we are actually converging toward. They are theatre. The thing already moving, already incentivised, already being deployed, is far more boring and far more concrete than a rogue mind: it is the steady removal of the leverage ordinary people have always held over the powerful.
Combine AI with decades of robotics, military technology, automation, sensors, surveillance infrastructure, compute concentration and platform economics, and the system-level implication is painfully simple. We are building the capacity to run more and more of society — and of warfare — with fewer and fewer humans. Not just fewer humans working. Fewer humans needed.
That single shift rewrites power dynamics more completely than any "AI goes rogue" storyline. Throughout history the elites — however you define them in a given era — always carried one constraint: they still needed the many. To produce food. To build. To transport. To fight. To enforce. To administer. To propagate beliefs. To keep the machine running. Automation reduces that dependency, AI accelerates the reduction, and interconnected infrastructure makes the reduction durable.
So the distribution of power is not one priority among many for anyone who cares about governance, resilience, freedom or the future of civilisation. It is essential. Even if you grant that the people in power today hold good intentions — a bold assumption — what happens when someone climbs the ladder and inherits a system offering unchecked, uncompromised, unquestioned power? Designing society around permanent benevolence is childish engineering.
There is one rule underneath all of it, and it is short enough to say in a breath. If the system needs you, you have leverage. If it doesn't, you don't. That is the whole engine. It explains why labour movements mattered when production was labour-intensive, why communities had leverage when local systems were necessary, why regimes fear strikes, defections, whistleblowers and refusal, and why dissent historically had at least some chance of creating friction.
If the system needs you, you have leverage. If it doesn't, you don't.
Now hold that rule against what is being built. Production systems that require fewer workers. Logistics systems that require fewer humans. Narrative systems that require fewer journalists and editors. Enforcement systems that require fewer officers willing to say "no". Discovery and innovation increasingly gated behind compute and capital. Read through the rule, the pattern is not a list of jobs disappearing. It is leverage being removed.
This is the distinction the jobs debate keeps missing. "Jobs being replaced" frames the loss as economic — a transition, a retraining problem, a question of new roles. Leverage being removed is a political fact. And once leverage is removed, participation becomes decorative. You can still act inside the system; your action simply no longer has to be answered.
Most people talk about power as if it were a moral flaw ("bad people"), a conspiracy ("shadowy cabals"), or a vibe ("the patriarchy", "the elites", "the system"). Power is usually more boring than any of that — and therefore more dangerous. It is structural. It lives in systems, not in villains. To see it clearly you need a map, and the map has four interlocking dimensions that societies have always cycled through.
Belief systems, education, culture, legitimacy, information control, attention allocation — the machinery that decides what counts as real and what is worth saying.
The coercive apparatus: enforcement, surveillance, military capacity, policing, deterrence, incarceration, sanctions — the capacity to compel.
Material foundations: water, food, energy, housing, transport and logistics, healthcare access, land, raw materials, employment systems — the things you cannot live without.
Scientific and technological capacity: AI, compute, chips, R&D, infrastructure, data, platforms, encryption and cyber, know-how — the capability to build and to discover.
Power concentrates when these four dimensions converge into a single point. Historically they were separate — different institutions held different pieces, and that separation produced cracks, frictions, competing elites, countervailing forces. Today that separation is collapsing. AI does not merely "improve productivity." It acts like a unified layer that can dissolve the remaining friction points between the four, binding them into one stack.
In older systems, anyone who wanted total power ran into messy bottlenecks. You could control the narrative, but you still needed farmers and workers. You could control production, but you still needed enforcers with families and consciences. You could control force, but you still needed legitimacy — or at least silence. You could control institutions, but you still depended on human competence spread across every layer. Each dependency was a negotiation point.
Now imagine those bottlenecks gone. Narrative shaped by a handful of algorithmic pipes. Essentials produced and distributed by automated supply chains. Enforcement increasingly remote, predictive and machine-mediated. Innovation gated behind compute infrastructure owned by a few entities. The friction points that used to force negotiation simply dissolve, one by one, until there is nothing left that the system has to bargain with.
That is not "a dictatorship." It can sit comfortably under democratic aesthetics — elections, courts, the full visible apparatus. It is simply a system where contestability collapses. And that is what absolute power looks like in the modern era: not a crown, not a palace, but an operating system.
People love the romantic idea that connectivity equals decentralisation. Sometimes it does. But in practice interconnectedness usually amplifies power rather than spreading it: it creates chokepoints (platforms, protocols, gateways, cloud infra), dependency (network effects, switching costs, lock-in), invisibility (you can no longer locate the decision-makers), and speed (central decisions propagate instantly). Technology does not invent domination — feudal lords, kings, empires and oligarchies got there long before. It makes domination more efficient, more scalable, and harder to pinpoint. It becomes unclear who is running the show, not because of a secret society, but because of a distributed structure of dependencies.
That is what "absolute power" looks like in the modern era: not a crown, not a palace — an operating system.
Take the four dimensions one at a time and watch the same move repeat: each had a human constraint built into it, and automation is quietly engineering that constraint out. This is not four separate stories. It is one mechanism applied four times.
The historical bargain of automation — that it frees humans from drudgery, raises productivity and creates new roles — has often been true. But it is not a law of nature. It is a political-economic arrangement. Automation "frees" people only if the gains translate into distributed ownership, wage growth, reduced hours, access to education and mobility, real opportunity, and a functioning social contract. What we are seeing in many places is the opposite: eroding job stability, hollowed-out middle roles, precarious work normalised, people competing for unstable income streams while the owners of automated systems capture the surplus. A system can be technologically advanced and socially regressive at once. So the question is not "will automation create new jobs." It is: will it keep the many structurally necessary to the few, or make the many surplus to the system? That is the moment the power-balance contract breaks — and when it breaks you do not get a clean transition, you get volatility, resentment, scapegoating, and stronger incentives for control.
This is where people get emotionally lazy. They imagine tyranny as a dictator, a coup, a sudden loss of freedom, obvious oppression — something you would recognise on sight and resist. Modern power is more subtle than that. It does not march in. It is installed, gratefully, by the people it will govern.
It arrives as convenience, as safety, as efficiency. It arrives wearing the language of "trust and safety" governance, "risk management", "national security", "fraud prevention", "harm reduction". Some of these are legitimate concerns — that is precisely why it works. The danger is not that the words are fake. The danger is that they become the moral cover for building systems that are unappealable, un-auditable, uncontestable, and structurally one-way.
Once systems become one-way, "democracy" becomes a branding layer.
If you want to test a sector against this, run a 2035 audit. Pick one domain you care about — health, education, energy, housing, finance, food, logistics, media, governance — and map it onto the four dimensions. Which narratives legitimise it? Which coercive tools enforce compliance? Which essential dependencies does it control? Which tech stack makes it possible? Then ask two questions: by 2035, which parts of this sector can be run with so much automation that fewer humans are structurally needed — and if fewer humans are needed, how do humans dispute decisions? Then the real one: who owns the automation, who governs it, what is the appeal mechanism, what is the exit option? If your answer is "there is no exit," you are looking at a power lock-in.
Power concentration is not always gradual. Sometimes it jumps. Three trigger patterns are realistic enough to deserve attention. Trigger one — AI supremacy inside a major institution: an "AI coup" without the drama. Not tanks in the street. A corporation or state reaches a capability lead and integrates it so deeply into governance, security, finance, production and narrative distribution that it becomes structurally unchallengeable. The coup is dependency. Trigger two — crisis capitalism on steroids: crises have always accelerated consolidation. When systems break, people accept centralisation "temporarily," assets get acquired cheaply, emergency measures become permanent, infrastructure is rebuilt under tighter ownership. AI makes crisis response faster — and crisis consolidation easier. Trigger three — full techno-autocracy becomes exportable: a state or alliance proves a model of total tech control that is stable, economically functional and effective at neutralising dissent, then exports it through security partnerships, infrastructure deals and platform ecosystems. Once that model is normalised as "modern governance," resistance gets framed as primitiveness.
None of this is destiny. Incentives push toward the worst equilibrium by default, but where we land depends on ownership and governance — and there are three plausible destinations.
Automation to abundance with agency. Essentials get cheaper, work becomes less coercive, ownership models distribute the gains, open standards reduce dependency, coercive automation is constrained by hard rules and auditability, and people retain real exit options.
Essentials are cheap but controlled. Labour loses leverage, narrative is engineered, enforcement is predictive and automated. People become managed, not represented.
Volatility increases, fear drives centralisation, and systems oscillate between breakdown and tighter control — each crisis ratcheting the grip a little further.
The asymmetry is the point. If we do nothing, we do not drift to Path A. We drift to whatever best serves existing concentrated incentives. Path A has to be built on purpose; Paths B and C are what you get by default. That is why power distribution should be the north star of systems change: if you do not treat it as a primary objective, your wins are temporary, because concentrated power can reverse almost anything — climate policy diluted, education turned into ideological conditioning, healthcare made extractive, media captured, institutions hollowed, reforms reduced to theatre. Solve downstream symptoms while upstream power keeps concentrating, and you are doing maintenance on a machine designed to re-break itself.
The point of the four-dimension chart is not to "understand power." It is to diagnose lock-in. Look at it through a 2035 lens and ask one question of each dimension: will it be automated enough that the number of humans required for contestability drops below a critical threshold? Because once contestability falls below that line, the game changes. You can still speak. You can still vote. You can still protest. But structurally, the system no longer has to negotiate with you. That is the line.
So forget the moral debates for a second and take one variable: how contestable is power? Can decisions be challenged? Can systems be audited? Can people exit? Are there competing providers? Are essentials resilient locally? Are enforcement mechanisms bounded? Is narrative distribution plural? Is knowledge production open enough to prevent dependency capture? If contestability is rising, you are building a civilisation with agency. If it is falling, you are building a civilisation with managed populations. AI can support either; the default incentives push toward the second.
This is why the counter-design has a shape. Reduce chokepoints — interoperability is not a technical preference, it is freedom infrastructure. Distribute ownership and control of essential systems, the part people avoid because it forces them out of abstraction. Put hard constraints on coercive automation: anything affecting rights, access, mobility, livelihood or freedom needs auditability, appeal mechanisms, due process, transparency around decision criteria, and strict limits on autonomous enforcement — because if enforcement becomes a black box, "justice" becomes an API call. Decentralise capability, not just access: "everyone can use AI" is still dependency if the capability is centrally owned. And measure power the way we measure carbon, because if you cannot measure concentration you cannot manage it — and we do not yet have mature power metrics, however much we pretend we do.
If contestability is rising, you're building a civilisation with agency. If contestability is falling, you're building a civilisation with managed populations.
None of this is an anti-AI argument. It is an anti-naivety one. AI and robotics can genuinely improve human life — but only if we stop pretending that "progress" automatically distributes itself. If we do not intentionally architect for distributed power, we drift toward an automated convergence that makes past empires look primitive, precisely because those empires still depended on humans in ways modern stacks increasingly will not. So yes, think about alignment. But also think about something more immediate: what happens when society becomes operationally runnable without most people — and the people who inherit the stack no longer need consent, labour, or even attention from the many? That is not a thought experiment for 2100. It is the direction of travel of the next decade, to an extent our civilisation has never faced before.
If the system needs you, you have leverage. If it doesn't, you don't.