Game Theory Warns AI Dependence May Trap Humanity in Comfortable Servitude

A provocative new game-theoretic analysis argues that the world’s headlong rush into artificial intelligence may be constructing a trap unlike any studied before: a collective dilemma in which every individual’s rational choice leads, inexorably, to a society that has surrendered its cognitive independence. Writing in the journal AI & Society, economist Taylor Collins of the University of the Incarnate Word proposes what he calls the Tragedy of Voluntary Servitude, a formal model showing that deep AI dependency can emerge as the unique stable outcome of a strategic game even when every participant would prefer that nobody become dependent at all.

The model borrows its skeleton from the most famous dilemma in the social sciences, the Tragedy of the Commons, first formalized by Garrett Hardin in 1968. In the classical version, a shared pasture or fishery is destroyed because each herder profits from adding one more animal, even though the aggregate result is ruin for all. Collins replaces the grass with something intangible: the distribution of cognitive and economic power across a population. Each player in his game faces a stark binary choice. Submit, and adopt AI as the primary engine of analysis, writing, planning, and decision-making. Or Resist, and maintain independent human capability at the cost of productivity.

The mathematical heart of the paper is a dominance proof. Each player’s payoff combines three components: productive capacity, which is high for submitters and low for resisters; competitive position, which rewards being augmented when others are not and punishes being unaugmented when others are; and autonomy, which erodes for everyone as more of the population submits and a central AI provider accumulates data, market power, and institutional influence. The critical asymmetry is that an individual’s own submission changes collective autonomy by only a vanishing fraction, an amount that shrinks as one over the population size. The productivity gain and competitive advantage, by contrast, are large and fully captured by the individual. Submit therefore strictly dominates Resist for any reasonably large population, and the unique Nash equilibrium is universal submission.

What makes the result unsettling is its robustness. Collins shows that the dominance of submission holds even if every player would individually prefer a world of universal resistance, because no single person’s restraint saves more than a negligible sliver of the commons while costing them the full productivity and competitive penalty. The collectively preferred outcome is individually inaccessible. The result also survives uncertainty about whether AI use actually erodes human cognition. The paper reviews striking empirical evidence that it may: endoscopists who regularly used AI-assisted polyp detection saw their unassisted detection rates fall from 28 to 22 percent within three months, and high school students given unrestricted access to GPT-4 performed 17 percent worse on subsequent unassisted mathematics exams than peers who never used the tool. But even if AI use were purely augmenting, Collins argues, the strategic dynamic toward dependency and power concentration would persist, because the mechanism is structural, not cognitive.

The paper then identifies six features that distinguish this new tragedy from the classical commons and make it harder to escape. The damage is invisible, because every step deeper into dependency arrives bundled with genuine benefits, so depletion feels like improvement. The lock-in is irreversible, since skills atrophy and alternative infrastructures wither, converting a repeated choice into a ratchet. There is no moment of crisis, no day the fish run out, and therefore no coordination signal to galvanize collective action. The subjects actively defend their own servitude, sincerely reporting that the tool has made them more capable, because the instrument they would use to evaluate their condition is the very instrument they have surrendered their judgment to. And the ultimate beneficiary of each submission is not the user but the central entity, which converts billions of individually rational choices into concentrated power.

That last point echoes a question posed in 1577 by Étienne de La Boétie, whose Discourse on Voluntary Servitude asked where the tyrant obtains the eyes to spy on citizens if the citizens do not provide them themselves. Collins formalizes La Boétie’s paradox: the AI provider’s power flows entirely upward from the population’s own choices, yet no individual can withhold their contribution without suffering a loss that dwarfs their marginal impact on the collective outcome. The servitude is voluntary not because people are ignorant of the risks, but because individual incentives overwhelm collective consequences regardless of what anyone knows.

Whether the resulting equilibrium constitutes a tragedy, Collins is careful to note, depends on the welfare framework applied. Under classical utilitarianism, universal submission may even be superior, since the augmented economy could produce unprecedented material abundance. Under the capability approach of Amartya Sen and Martha Nussbaum, cognitive autonomy is constitutive of meaningful agency and cannot be freely traded for productivity. Under the republican tradition of freedom as non-domination, a society whose cognitive infrastructure is controlled by an entity accountable chiefly to its own interests is unfree regardless of how prosperous its members are. The strategic dynamics, however, are indifferent to this philosophical debate; the equilibrium arrives either way.

The paper sketches what a society at the universal submission equilibrium might look like. Material wealth could be historically unprecedented, with diagnostics, legal analysis, and creative production abundant and cheap. But cognitive homogeneity would replace intellectual diversity, as everyone thinks through the same system trained on the same data, narrowing the range of ideas considered and leaving the society capable of optimizing within existing frameworks but unable to generate new ones. Democratic governance would be structurally undermined, since citizens relying on the same AI to understand policy and evaluate candidates would hand its operator an influence over public deliberation that no propaganda apparatus has ever achieved. A new class division would separate those who own the infrastructure from those who depend on it, and the servitude would be stable precisely because it is comfortable.

Can the trap be escaped? Collins audits Elinor Ostrom’s celebrated eight design principles for governing commons and finds the fit sobering. Only two, congruence with local conditions and nested enterprises, apply readily. Boundary definition and monitoring face severe challenges, because the resource being depleted is internal and invisible. Graduated sanctions must be inverted, replacing punishment for overuse with incentives for demonstrated independence, such as professional benchmarks of unassisted performance. The paper also surveys antitrust tools, including interoperability mandates and public compute options, open-source ecosystems, algorithmic auditing under frameworks like the EU AI Act, and education designed to cultivate the capacity to think without assistance. Cultural norms, Collins argues, are the only mechanism that changes the payoff structure itself, but they evolve on generational timescales while AI adoption moves on quarterly ones.

The author is explicit about the model’s limits: it assumes a binary choice, homogeneous players, a static one-shot decision, and linear autonomy dynamics, and it presents no new empirical data. He also acknowledges, with some irony, that the paper was developed through extensive collaboration with a large language model, a choice he describes as perfectly rational under the very incentive structure the model describes. Yet the central warning stands independent of those simplifications. The problem, Collins concludes, cannot be solved by individual restraint, because dominant strategies are too powerful; it demands institutional, structural, and system-wide responses, built before lock-in completes. If La Boétie was right that servitude endures through the atrophy of the imagination of alternatives, then waiting until the dependency is complete to act may not merely be difficult. It may, by that point, be literally unthinkable.

Subject of Research: A game-theoretic model of cognitive dependency on artificial intelligence as a collective-action tragedy

Article Title: The Tragedy of Voluntary Servitude: a game-theoretic model of cognitive dependency in the age of artificial intelligence

Article References: Collins, T. (2026). The Tragedy of Voluntary Servitude: a game-theoretic model of cognitive dependency in the age of artificial intelligence. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03309-x

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03309-x

Keywords: game theory, artificial intelligence, cognitive dependency, tragedy of the commons, voluntary servitude, Nash equilibrium, power concentration, cognitive deskilling, large language models, collective action, AI governance, autonomy