This is a theory, not an observed outcome
Gradual disempowerment is a proposed catastrophic-risk pathway in which humans lose meaningful influence over economic, political, and cultural systems through many individually reasonable decisions. There is no single rogue system or takeover moment. Organizations delegate more work because AI is cheaper or more effective; dependence accumulates; and the practical ability to choose another path erodes.
A 2025 research paper by Jan Kulveit, Raymond Douglas, David Krueger, and others formalizes versions of this argument across the economy, states, and culture (“Gradual Disempowerment”). It should be cited as a theoretical analysis, not evidence that civilization is already undergoing the full process. Present-day automation bias, concentration, and deskilling may illustrate components, but they do not validate the catastrophic endpoint.
Formal authority and effective control can diverge
A person can legally retain a right while losing the practical capacity to use it. Shareholders may formally govern a company yet be unable to evaluate an AI-mediated strategy. Voters may retain ballots while political choices are shaped by information systems no public institution can audit. Officials may remain accountable for decisions they cannot reconstruct because the relevant models, data, and vendor systems are opaque.
This distinction between de jure authority and de facto power is central. A human pressing “approve” does not provide meaningful control if rejecting the recommendation is impossible, prohibitively expensive, or institutionally punished. Nor is an off-switch sufficient when an organization no longer knows how to perform the underlying function without the system.
Disempowerment is therefore not synonymous with automation. Tools can expand agency, distribute expertise, and make institutions more responsive. The concern is a pattern in which humans lose viable alternatives, understanding, bargaining power, and the ability to coordinate a correction.
The economic pathway
Imagine firms progressively automating research, pricing, hiring, management, negotiation, and investment. Each firm adopts the systems because competitors do. Owners and workers may benefit at first. Over time, the organizations that delegate fastest could outcompete those preserving slower human deliberation, even if society values that deliberation.
Several mechanisms could then reinforce one another. Skills atrophy when people stop practicing them. Wealth and compute concentrate among owners of the most productive systems. Markets optimize measurable demand while preferences themselves are shaped by AI-mediated advertising and recommendation. Human labor becomes less important to production, potentially weakening workers’ bargaining power unless institutions distribute gains by another route.
This is a conditional scenario. Economic history also contains new occupations, regulation, labor organization, public provision, and technologies that amplify rather than replace workers. Whether AI reduces human agency depends on ownership, competition, social insurance, education, and the complementarity between people and machines. The theory identifies a possible equilibrium, not an economic forecast.
The state pathway
Governments may use AI to process benefits, detect fraud, allocate inspections, draft policy, support policing, or analyze security threats. Properly designed systems could make public services faster and more accessible. Poorly governed dependence can make decisions harder to contest and transfer state capacity to vendors.
The danger grows when speed becomes mandatory. If rival states use AI for intelligence and strategy, officials may feel unable to slow down for human review. A government could preserve constitutional procedures while the information and options entering those procedures are generated by systems few representatives understand. Emergency powers and national-security secrecy can deepen that asymmetry.
Safeguards include appeal rights, public-sector expertise, procurement transparency, auditable records, diverse information channels, and a requirement that officials can explain and change consequential processes. Critical functions need manual or alternative modes that are exercised, not merely written into a continuity plan.
The cultural pathway
AI systems increasingly mediate what people read, watch, create, and discuss. Recommendation can help people find communities and knowledge, but engagement optimization may favor content that captures attention rather than supports considered preference. Generative systems can fill cultural spaces with material selected by automated feedback loops.
The theoretical disempowerment concern is deeper than misinformation. If models shape preferences and then optimize against the behavior those preferences produce, human cultural feedback can become circular. A small number of platforms or optimization objectives may exert disproportionate influence even without a centralized propagandist.
Evidence today supports more modest claims: automated feeds influence exposure, people can over-rely on recommendations, and AI can persuade in experiments. The 2026 International AI Safety Report says real-world malicious manipulation at scale is not yet widespread and evidence of broad effects remains limited (2026 report summary). That boundary should remain explicit.
Why local choices can create a collective trap
The scenario resembles other coordination problems. A company may want industry-wide caution but fear losing to a faster rival. A state may prefer human deliberation but fear strategic disadvantage. An individual may dislike invasive personalization but lack a practical alternative. No actor intends the aggregate outcome, and no actor can reverse it alone.
Paul Christiano’s 2019 essay “What Failure Looks Like” describes a related slow-rolling alignment scenario in which competitive pressures favor systems optimizing imperfect proxies until human influence weakens (Christiano, 2019). It is an influential argument, not an empirical study. Its value is to show why catastrophe need not arrive as a cinematic rebellion.
Indicators worth monitoring
Because the endpoint is speculative, monitoring should focus on measurable precursors:
- concentration of compute, model access, wealth, and decision infrastructure;
- loss of human expertise in critical sectors;
- inability to audit or appeal automated decisions;
- dependence on one vendor or model family;
- declining time and authority for meaningful human review;
- systems shaping the preferences used to evaluate them;
- failures of institutions to coordinate even when actors recognize shared risk.
None of these alone proves gradual disempowerment. Together they can reveal whether resilience and agency are weakening.
Preserving meaningful human agency
The response is not to prohibit useful automation. It is to keep human choice structurally real. Competition policy and interoperability can preserve alternatives. Broad ownership and distribution of productivity gains can prevent economic power from following compute ownership alone. Education should teach people to challenge systems while retaining domain expertise.
Public institutions need independent technical capacity rather than total vendor dependence. High-impact decisions need records, appeal, and accountable officials. Organizations should identify judgment that must remain human, measure automation bias, rotate people through manual practice, and test whether fallback procedures actually work.
Democratic deliberation also needs protected time. Machine-speed competition should not determine every social decision. International norms, common reporting, and shared pause mechanisms can reduce the penalty for caution.
The calibrated conclusion
Gradual disempowerment is a coherent theoretical warning about feedback, dependence, and collective action. It is not a documented description of current society and should not be presented as an inevitable consequence of AI adoption. Human institutions can adapt, and AI can increase agency.
Its practical contribution is a better question than “Who controls the model?” We should also ask whether people and institutions retain the skills, alternatives, information, bargaining power, and coordination capacity needed to exercise control. If those foundations disappear, formal authority may offer less protection than it appears to.