3. The Multiplayer Prisoner’s Dilemma

The previous two essays established the technical and propagation conditions. Unit cost dominance has been crossed for a substantial fraction of professional cognitive tasks. Interface collapse is the propagation mechanism that carries task-level dominance into workflow recomposition and labour-market non-absorption. The question this essay addresses is what follows once both conditions hold. Specifically, can the deployment of AI be slowed, paused, or restrained by any actor or coalition of actors? The answer is no, not because no actor wishes to restrain it, but because the structure of competition makes restraint a dominated strategy at every level at which it could be attempted.

The argument is not that coordination is impossible in principle. It is that restraint is non-pausable as a competitive equilibrium under the conditions that obtain. The action of restraint is physically available. It is dominated by the action of deployment at every level. The actors who choose restraint are eliminated by the actors who choose deployment.

This is the political-economy hinge of the thesis. Without it, unit cost dominance and interface collapse would be problems amenable to coordinated response. With it, they become structural inevitability.

The fractal structure

The classical Prisoner’s Dilemma involves two actors choosing between cooperation and defection, with payoffs structured so that defection dominates regardless of what the other actor does. Both actors defect even though both would prefer mutual cooperation. The lesson is that individual rationality can produce collective outcomes no one wanted.

The AI deployment situation is a fractal version of the same structure. The same payoff shape repeats at four levels: worker, firm, sector, and state. At each level, the actor who deploys gains a relative advantage. The actor who restrains is competitively dominated. Restraint is not punished by some external force. It is punished by the actors who do not restrain. The dilemma is recursive, and it operates simultaneously at every scale.

The fractal structure is what makes the situation unrecoverable through any single intervention. A regime that addresses the firm level is undermined by worker-level adoption. A regime that addresses the worker level is undermined by firm-level deployment. A regime that addresses the state level is undermined by both. The layers reinforce each other. Defection at any layer creates pressure for defection at the others.

The worker layer

A knowledge worker faces the dilemma directly. They are not making a grand ideological choice. They are trying to stay employed.

If the worker uses AI tools, they produce more, learn the tools faster, become cheaper per unit of output, and survive the next round of cuts. They also train themselves on the systems that will eventually replace their entire role, but that loss is years away and the layoff is months away. The discount rate makes adoption rational.

If the worker abstains, the colleague who uses AI produces more, learns faster, costs less per unit of output, and gets the role the abstaining worker wanted. Abstention is punished immediately. The collective benefit of abstention, if any, accrues to a future labour market the abstaining worker may not be in.

This matters because it makes the whole system harder to govern than the firm-level analysis suggests. Even if firms signed restraint agreements, workers would still adopt AI inside workflows. Even if states regulated formal deployment, individuals would still use tools for drafting, coding, analysis, planning, search, research, summarisation, and verification. Adoption happens below the level at which coordination can see it. The worker layer is the bedrock of the dilemma. Any regime that fails to account for worker-level adoption has already failed as wage-demand circuit defence.

The firm layer

A firm faces the same payoff shape one level up. The firm that deploys AI plus verification reduces unit costs for affected cognitive tasks. The firm that does not deploy faces competitors with lower costs. The undeploying firm either matches the cost reduction (which requires deployment) or loses market share until it exits. There is no third option.

This is true even if every executive at every firm prefers to preserve human employment. Their preferences are dominated by the competitive structure. In contestable markets, over sufficient time, the firm that persistently prioritises employment over unit cost is selected against. The firm that survives is the firm that deploys. Boards and executives are evaluated on quarterly performance. They cannot maintain restraint in the face of competitor deployment without losing their positions. Any individual firm that tries to restrain deployment is replaced by leadership that will deploy.

The second-best model in any given category may survive in niches, but it does not set the platform terms. The firm that is competitively second-best in its category does not face zero revenue. It faces shrinking margins, weaker pricing power, and slower talent acquisition. None of these is fatal in the short run. All of them are corrosive in the medium run. The firm has to keep matching the leader to avoid drift toward irrelevance.

The Dollar Auction layer

In a normal Prisoner’s Dilemma, defection explains why actors deploy. It does not fully explain why they continue spending after returns become uncertain. For that, the AI race needs a second mechanism: Dollar Auction dynamics.

In the classic Dollar Auction, the winner receives the dollar, but the losing bidder still pays their bid. This creates escalation. Once both players have bid heavily, quitting guarantees loss. Continuing is irrational from the outside but rational from the player’s position inside the game. The game is structurally an all-pay tournament.

AI infrastructure has this structure. The losing firms do not simply fail to win the prize. They are left with stranded compute, depreciating models, lost talent, weakened distribution, reduced investor confidence, and no claim on future AI rents. The winner does not merely earn revenue. The winner sets the platform layer on which others build.

This is why continued spending can remain rational even when near-term margins are poor. The sunk cost is not the reason. The option value is. Stopping destroys the option. Continuing preserves a non-zero chance of surviving the next round. The frontier labs are not spending hundreds of billions because they are confident of the return. They are spending because withdrawal would foreclose the position from which any future return could be claimed.

The race is therefore not just a Prisoner’s Dilemma. It is a Prisoner’s Dilemma with all-pay tournament dynamics. Everyone pays to stay in the game. Losing does not mean returning to the starting line. Losing means paying for the race and receiving no platform position.

The sector layer

At the sector level, restraint is unstable because capital, talent, and customers migrate toward lower-cost substitutes and adjacent entrants. A restrained sector does not compete only with its existing incumbents. It competes with new firms, foreign firms, platform firms, and adjacent sectors that use AI to repackage or automate part of its function. The restrained sector becomes a protected high-cost island. Unless the protection is global and permanent, the island erodes.

A restrained legal services sector does not lose customers to manufacturing. It loses customers to legal-tech platforms, to in-house counsel teams equipped with AI tools, to foreign legal services providers operating remotely, and to adjacent service categories that absorb part of what was previously legal work. A restrained creative services sector loses to platforms, to in-house creative teams, to foreign agencies, and to direct AI-assisted production by clients. The pattern repeats across every cognitive-services sector.

Industry-level coordination is theoretically achievable. It is also structurally undermined by the boundary between industries being more permeable than the coordination assumes. Sector boundaries are administrative conveniences, not economic walls. AI dissolves the conveniences without asking anyone’s permission.

The state layer

A state that restrains AI deployment loses ground in technology, in productivity, in fiscal capacity, and in the strategic capabilities that increasingly depend on AI infrastructure. Other states do not restrain. The restraining state’s relative position weakens. Its capacity to fund any social policy, including the social policy that might compensate displaced workers, declines. The restraining state arrives at a worse position for its own population than the state that deployed.

State-level restraint, even if politically achievable in any single state, produces worse outcomes for that state’s population than deployment does. The political coalitions that supported restraint weaken as the costs become visible. Capital flight, brain drain, and erosion of competitive position appear within years. The state either abandons restraint or accepts persistent relative decline.

National security amplifies this dynamic. AI is increasingly framed as military, intelligence, scientific, industrial, and cyber capacity. A state that falls behind in AI falls behind in those capacities. No major state can accept that outcome, regardless of the social cost of deployment, because the alternative is being out-positioned by states that did not accept it.

Friction changes timing, not equilibrium

The standard objection to this analysis is that real deployment is slow. Firms face integration costs, liability concerns, regulatory delay, internal politics, union resistance, customer scepticism, professional gatekeeping, data privacy obligations, and infrastructure constraints. The objection grants the competitive logic but argues that friction prevents the logic from operating at the speed the thesis implies.

Friction is real. The thesis does not deny it. The thesis denies that friction constitutes a stable cooperative equilibrium. A firm may delay because of liability, culture, integration cost, or regulatory uncertainty. If unit cost dominance holds and competitors deploy, the delay becomes a cost disadvantage. The firm that delays still has to deploy eventually, or accept persistent erosion of competitive position. Friction redistributes when the deployment happens. It does not change whether it happens.

The same applies at every layer. Workers who delay adoption are outcompeted by workers who do not. Sectors that delay are eroded by adjacent sectors and foreign entrants that do not. States that delay accept relative decline. In each case, friction is a transient cost, not a permanent equilibrium. The friction-as-rescue argument requires friction to be persistent and universal, which it is not, because the actors who absorb the friction are competing with actors who do not.

This is why drag is not rescue. The closing essay in this sequence develops the point in detail. The point here is narrower. Within the Multiplayer Prisoner’s Dilemma, friction modulates timing. It does not produce the cooperative equilibrium that would be required to restrain deployment.

Why coordination is not an equilibrium

The standard answer to a Prisoner’s Dilemma is coordination. If the actors can communicate, commit, and enforce, they can reach the cooperative outcome that both prefer. International treaties, industry standards, and binding agreements are the institutional forms this coordination takes. The question is whether any of them can stabilise restraint in the AI case.

Repeated interaction exists. The classical route out of a Prisoner’s Dilemma requires not merely repeated interaction but observable defection and credible enforcement. Those are the missing conditions. In AI deployment, defection is internal, diffuse, continuous, and often indistinguishable from ordinary productivity improvement. The actors capable of enforcing restraint are also the actors most exposed to the cost of restraint.

Observable defection breaks down because AI deployment produces no outward sign. The output is the same. The product is the same. The customer interaction is the same. Detecting that the production process has changed requires inspection of internal workflows, which firms have strong incentives to obscure and which regulators lack the capacity to perform at scale. Aggregate data, payroll records, sectoral productivity statistics, and wage patterns can detect deployment after the fact. By the time the data shows defection, the defector has already captured the competitive advantage. Macro-level observability is post-hoc. Treaty enforcement requires real-time observability or it is not enforcement.

Credible enforcement breaks down because the actors with the capacity to enforce are the same actors with the strongest incentive to defect. States are the enforcers of international agreements. States are also the actors competing for AI advantage. A state that genuinely enforces a restraint agreement against its own firms cedes competitive position to states that do not. The enforcement function and the defection incentive are colocated in the same institution, which means enforcement is structurally undermined by the same dynamic that produced the original problem.

The conclusion is not that coordination is logically impossible. It is that coordination is not an equilibrium under conditions of weak enforcement, high strategic uncertainty, and continuous technological change. Those conditions are not contingent features of the present moment. They are structural features of the AI development environment.

Why the precedents do not transfer

Defenders of AI restraint sometimes point to historical precedents for coordinated restraint of dangerous technologies. Nuclear non-proliferation. Chemical weapons treaties. Ozone-depleting substances. Each is invoked as evidence that coordination is possible. The evidence is real but does not generalise to AI for reasons worth being precise about.

Nuclear non-proliferation works as well as it does because the core pathway to the weapon runs through scarce fissile material and identifiable facilities. AI has some chokepoints too: advanced chips, hyperscale data centres, cloud providers, and energy infrastructure. Those chokepoints matter for frontier training. They matter less for the wage-demand problem. Labour substitution diffuses through APIs, open-weight models, smaller specialised systems, enterprise wrappers, and ordinary workflow tools. A regime that slows frontier training does not necessarily stop cognitive substitution below the frontier.

Chemical weapons treaties work better than AI restraint because chemical weapons have little legitimate civilian economic value. Producing chemical weapons does not generate revenue, market share, or competitive advantage in any non-military domain. States that abandon chemical weapons production lose nothing economically. AI, by contrast, is the most economically valuable technology of the era. States that abandon AI development lose enormous economic value. The cost-benefit calculation that supports chemical weapons restraint does not apply to AI restraint.

Ozone-depleting substances were restrained because the relevant chemicals had close substitutes that did not damage the ozone layer. Industry could shift to substitutes without losing economic function. AI has no substitute for the cognitive functions it performs. The substitute for AI cognition is human cognition, which is what AI cognition replaces because it is more expensive. The Montreal Protocol model assumed an off-ramp that does not exist for AI.

The AI race is not a Stag Hunt because there is no shared stag. Third-party defection by open-weight developers, foreign states, market entrants, or internal workflow adopters destroys the cooperative equilibrium. It is not a nuclear arms race in the ordinary sense because AI is not only a weapon. It is a productive asset. A state that restrains nuclear weapons may preserve most of its economy. A state that restrains AI restrains its own productivity growth, fiscal capacity, and strategic infrastructure. It is closest to Moloch, but Moloch alone is too static. The AI race has a burn mechanism. It is Moloch with a balance sheet.

The precedents that critics invoke do not show that AI restraint is possible. They show that restraint is possible under specific conditions that do not obtain in the AI case. The relevant question is not whether some technologies have been restrained. It is whether the conditions that allowed those restraints are present here. They are not.

The recursive prize

A normal Dollar Auction has a fixed prize. The game ends when someone wins the dollar.

The AI race does not have that structure. The prize is recursive. Each capability level unlocks the next capability level. Each deployment produces revenue, feedback, workflow data, product telemetry, distribution, and political leverage that can be used in the next round. Winning one round does not end the contest. It changes the baseline from which the next contest begins.

This matters because exit calculation becomes unstable. In a finite game, a firm can model the endpoint and decide whether the remaining prize justifies the remaining spend. In the AI race, there is no final prize against which exit can be calculated. The actor that exits does not preserve its position. It falls behind the moving frontier.

You do not win the AI race. You survive one round. Then the next round starts immediately, with higher stakes and faster timelines. There is no plateau. Only slope.

The race is therefore not won once. It is survived repeatedly. The absence of a finish line does not free the players from the race. It removes the only thing that could have stopped it.

Bridge to Sorites

There is one more reason coordination fails, and it requires its own essay. Even if the institutional conditions for restraint were satisfied, the boundary between AI assistance and AI replacement could not be cleanly drawn. The next essay handles that problem. The Sorites Collapse Principle is what makes the Prisoner’s Dilemma irrecoverable. In the absence of Sorites, a sufficiently determined coordination effort might find an enforceable definition of what to restrain. With Sorites, no such definition exists.