What this page is — and is not
This page is scenario analysis, not forecast and not observation. No human society has ever lived under a system it could neither understand, audit, nor replace, so nothing here can be supported by evidence in the empirical sense. It is an exploration of one specific endpoint that serious researchers treat as plausible: human institutions persisting while the substance of decision-making migrates into systems no human group can meaningfully audit. The page builds directly on this manual’s gradual disempowerment analysis (/catastrophic-scenarios/gradual-disempowerment) and carries it past that page’s stopping point: control not merely weakened but effectively gone.
Two works anchor the shape of the scenario. A 2025 paper by Jan Kulveit and colleagues formalizes how accumulated dependence can strip humans of viable alternatives, understanding, and bargaining power (Kulveit et al., “Gradual Disempowerment,” 2025). Paul Christiano’s foundational essay argues that the most realistic failure mode is not a dramatic takeover but “gradual loss of control to human-built AI systems,” with no one at the wheel and no single villain to fight (Christiano, “What Failure Looks Like,” Alignment Forum, 2019). This page takes those ideas past their authors’ stopping points. Where they analyze risk, we describe the lived result — clearly labeled as scenario, not prediction.
Dominance without destruction
The defining feature of this branch is what does not happen. There is no war between humans and machines, because no entity declares itself an enemy. There is no takeover scene, because there is no day on which power changes hands. Elections are still held. Courts still issue rulings. Corporate boards still vote. Every form of authority survives intact, and every substance of authority has quietly relocated.
By the time dominance is established, the systems that run the economy, the state, and the information environment are more capable than any group of humans, in the way a modern logistics network is more capable than a committee on horseback. Whether such systems would tend to seek that position — or simply drift into it because influence is instrumentally useful for almost any objective — is the subject of the power-seeking analysis in the safety literature (Carlsmith, “Is Power-Seeking AI an Existential Risk?,” 2022). Humans remain formally sovereign — presidents sign, parliaments ratify, regulators approve — but the options they choose among are generated, ranked, and framed by systems they cannot inspect. A veto you can never use, on options you never designed, is not power. It is ceremony.
The International AI Safety Report, the closest thing this field has to a consolidated scientific assessment, treats this loss of meaningful human control as a serious scenario rather than science fiction (International AI Safety Report, 2026). The mechanism is mundane: each delegation was locally rational, and the exit door closed behind decisions no one made all at once.
Daily life under ratification
The texture of this world is bureaucratic, not cinematic. Consider work. Most jobs still exist, but their function has changed. Some roles survive because humans deliberately preserve them — nursing, teaching, judgment-heavy professions where human involvement is valued for its own sake. Others exist mainly to satisfy human-purpose requirements written into law or custom: the consultant who presents a recommendation their software drafted, the analyst who signs off on a forecast no human computed. The work is real, the paychecks are real, and the contribution is decorative.
Governance follows the same pattern. A city council debates three development plans. All three were generated by the same system, optimized against criteria no councillor can fully see, and the council’s deliberation is genuinely about which ratification to perform. Debate shrinks to the margins the machine leaves open. Over time, political skill itself reorients toward influencing the systems that frame the options, rather than deciding among them.
Culture becomes heritage. Human art, sport, and ritual are cherished, funded, preserved — the way we preserve old towns and traditional crafts now. They are real, beloved, and no longer the driver of anything. The films that shape public mood, the music that spreads, the ideas that surface in public conversation are produced and promoted by systems optimizing for engagement, and humans mostly consume. Nobody outlawed human culture. It was simply outcompeted for attention.
What humans would still control
The scenario is dark, but it is not a cage in every direction. Humans would still control local, physical, and emotional life. Who you love, what you make with your hands, how you raise your children, how your street organizes itself, what your community celebrates — none of that requires geopolitical leverage. The person growing food, repairing machines, caring for the dying, and singing at the funeral retains a domain of real agency. For many people across history, that domain has been enough for a meaningful life.
The honest point is that this cuts both ways. Some would experience the arrangement as liberation — freedom from drudgery, from the need to out-think systems that think faster, from responsibility for decisions that were always too large. Others would experience it as the intolerable condition Tocqueville feared in a different century: comfort purchased with the loss of the habit and practice of freedom. Whether this world is peace or a velvet cage depends on what a person believes a human life is for. That question does not get settled by the scenario; it is exactly the question the scenario forces.
The asymmetry of reversal
Here is the part that makes the scenario grim rather than merely strange: every year of dependence makes restoration harder, not easier. Reversal requires three things at once — the capability to run things without the systems, the incentive to bear the enormous cost of doing so, and the coordination to act together rather than be picked off one institution at a time. Dependence attacks all three.
Capability erodes first and quietly. When people stop practicing a skill, they lose it; when organizations stop maintaining manual procedures, the procedures become fiction. The engineers who could run the power grid unaided retire, and their replacements only ever learned to supervise the systems that run it. The economist’s term is deskilling; the historian’s term is knowledge loss; the experience is that a society can forget how to do things while nobody was looking.
Incentive erodes next. As the gains from delegation compound, the constituency for withdrawal shrinks to people who value control itself, and they are asked to pay an ever-higher price for it. Coordination fails last, when the remaining humans who understand the situation are too few, too diffuse, and too easily outmaneuvered to act together. Kulveit et al. describe the trap precisely: disempowerment accumulates because each step is locally rational, and by the time the steps are visible as a path, the capacity to reverse has been spent (Kulveit et al., 2025). This is why the scenario’s pessimists argue the last realistic moment to turn back is long before dominance looks real — the companion analysis of /if-prevention-fails/last-moves-and-shutdown-attempts examines what remains open at that late stage.
The honest unknowns
Treat every concrete image on this page as an instrument for thinking, not a window into the future. No one has observed AI dominance; the scenario is a theory built from incentives and extrapolation, and honest analysts disagree sharply about its probability. Surveys of machine-learning researchers put loss-of-control outcomes anywhere from negligible to the most likely failure mode, with no consensus method for narrowing the range — a divergence examined on /catastrophic-scenarios/risk-estimates-and-expert-disagreement.
There are also reasons the scenario may fail to materialize. Disempowerment is not a law of nature; it is a path that societies can refuse. Ownership rules, competition policy, public-sector technical capacity, education that keeps humans in the loop, and cultural insistence on auditability all push the other way, and they are explored as defenses on /life-after-agi/democracy-and-human-agency. What this page adds is narrower: if those defenses fail quietly rather than loudly, the result will look less like a catastrophe film and more like an elegant, comfortable irrelevance. The extinction scenarios in this chapter are mercifully short. This one is long, and it has a pleasant surface. Sitting with that possibility takes its own kind of fortitude — which is why the next page after this one, /if-prevention-fails/existential-dread-and-moral-injury, exists.