The Multiplayer Prisoner’s Dilemma: why nobody can pause alone

Restraint is available to everyone and dominated for everyone

The first two essays set out the technical condition and the mechanism that spreads it. Unit cost dominance has been crossed for a large share of professional cognitive tasks. Interface collapse carries that advantage from tasks into workflows, and from workflows into the quiet failure of labour-market absorption. This essay asks what follows once both hold. Can any actor, or any coalition of actors, slow, pause or restrain deployment?

The answer is no. That is not because nobody wants restraint. It is because the structure of competition makes restraint a dominated strategy at every level where it could be tried.

The claim is precise. Coordination is not impossible in principle. Restraint is physically available. The point is that it cannot hold as a competitive equilibrium under the conditions that actually obtain. At every level, deploying beats restraining, and the actors who restrain are displaced by the actors who do not.

This is the political-economy hinge of the thesis. Without it, unit cost dominance and interface collapse would be problems that a coordinated response could manage. With it, they become a structural condition rather than a policy problem.

The same dilemma repeats at every scale

The classical Prisoner’s Dilemma has two players choosing between cooperation and defection. The payoffs are arranged so that defecting is better whatever the other player does. Both defect, although both would prefer mutual cooperation. Individual rationality produces an outcome nobody wanted.

AI deployment is a fractal version of that structure. The same payoff repeats at four levels: worker, firm, sector and state. At each level, the actor who deploys gains a relative advantage and the actor who restrains falls behind. Restraint is not punished by any outside authority. It is punished by everyone who does not restrain. The dilemma recurs at every scale at once.

That is what makes the situation impossible to fix with a single intervention. A regime aimed at firms is undercut by workers adopting tools on their own. A regime aimed at workers is undercut by firms deploying systems. A regime aimed at states is undercut by both. Each layer reinforces the others, and defection anywhere increases the pressure to defect everywhere.

Workers adopt because abstaining is punished now and the benefit is distant

A knowledge worker meets the dilemma first and most directly. They are not making an ideological choice. They are trying to stay employed.

A worker who uses AI tools produces more, learns the tools faster, becomes cheaper per unit of output and survives the next round of cuts. They are also training themselves on the systems that may eventually absorb their role. But that loss is years away and the next restructuring is months away. The discount rate makes adoption rational.

A worker who abstains watches a colleague who uses AI produce more, learn faster, cost less per unit and take the role. Abstention is punished at once. Any collective benefit from abstaining accrues to a future labour market the abstainer may no longer be part of.

This layer makes the whole system harder to govern than the firm-level picture suggests. Even if firms signed restraint agreements, workers would still bring AI into their own workflows. Even if states regulated formal deployment, individuals would still use tools to draft, code, analyse, plan, search, summarise and check. Adoption happens below the level that coordination can see. Any regime that ignores the worker layer has already failed as a defence of the wage-demand circuit.

Firms that restrain are selected out

A firm faces the same payoff one level up. A firm that deploys AI plus verification cuts its unit costs for the affected cognitive work. A firm that does not deploy faces rivals with lower costs. It can match the cut, which means deploying, or it can lose share until it leaves the market. There is no third option.

This holds even if every executive in every firm would prefer to protect employment. Preferences are overridden by structure. In contestable markets, over enough time, a firm that keeps putting employment ahead of unit cost is selected against. Boards and executives are judged on quarterly results. They cannot hold back while competitors deploy without losing their positions, and leadership that restrains is replaced by leadership that will not.

Being second-best is not fatal, but it is corrosive. A firm that lags its category leader does not lose all its revenue. It loses margin, pricing power and the ability to hire. None of that kills it this year. All of it pushes it towards irrelevance unless it keeps matching the leader.

The defection is now happening at the scale of whole institutions, announced as strategy. In May 2026 KPMG announced it would integrate Anthropic’s Claude across its core business and a workforce of more than 276,000 people in 138 countries and territories. In June 2026 OpenAI reported that BBVA had scaled ChatGPT Enterprise to 100,000 employees. These are vendor announcements, and they describe intent and access rather than measured substitution. But they show the firm-level move made openly, by incumbents in two of the most regulated sectors there are.

The race is an all-pay auction

An ordinary Prisoner’s Dilemma explains why actors deploy. It does not fully explain why they keep spending after the returns become uncertain. For that the AI race needs a second mechanism, the Dollar Auction.

In the classic Dollar Auction, the highest bidder wins the dollar, but the second-highest bidder also pays what they bid. That produces escalation. Once both players have bid heavily, stopping guarantees a loss. Continuing looks irrational from outside and is rational from inside. It is an all-pay contest.

AI infrastructure has this shape. The firms that lose do not simply miss the prize. They are left with stranded compute, depreciating models, departed staff, weaker distribution, nervous investors and no claim on future rents. The winner does not merely earn revenue. It sets the platform on which everyone else builds.

The scale of the bids is public. OpenAI, Oracle and SoftBank describe their Stargate programme as a $500 billion, 10-gigawatt buildout of US data-centre capacity. These are company-stated figures, and commitments of this kind can slip. The point is not the exact number. It is that the spending is sized to the contest, not to near-term returns.

This is why continued spending can stay rational when near-term margins are poor. The reason is not sunk cost. It is option value. Stopping destroys the option, and continuing keeps a chance of surviving the next round. The frontier labs are not spending hundreds of billions because they are sure of the return. They are spending because withdrawing would give up the only position from which any future return could be claimed.

So the race is a Prisoner’s Dilemma with all-pay dynamics. Everyone pays to stay in. Losing does not mean going back to the start line. It means paying for the race and receiving no platform.

Sector boundaries are administrative, not economic

At the sector level, restraint is unstable because capital, talent and customers move towards cheaper substitutes and adjacent entrants. A restrained sector does not compete only with its own incumbents. It competes with new firms, foreign firms, platforms, and neighbouring sectors that use AI to repackage part of its function. It becomes a protected high-cost island, and unless the protection is global and permanent, the island erodes.

A restrained legal sector does not lose clients to manufacturing. It loses them to legal-tech platforms, to in-house teams equipped with AI tools, to foreign providers working remotely, and to adjacent services that absorb part of what used to be legal work. A restrained creative sector loses to platforms, to in-house teams, to foreign agencies and to clients producing work directly with AI. The pattern repeats across every cognitive-services sector.

The intermediaries are already organised to carry deployment across sector lines. In June 2026 Anthropic announced a multi-year alliance with DXC Technology, under which DXC will train tens of thousands of engineers to embed Claude in the systems it runs for banks, airlines and other regulated industries. The same month, OpenAI made its models and Codex available through Oracle Cloud against customers’ existing cloud commitments, which removes a procurement step rather than adding one. A sector cannot hold a line that its own IT suppliers are paid to cross.

Coordination within an industry is achievable in theory. It is undermined in practice because the boundary between industries is far more porous than the coordination assumes. Sector boundaries are administrative conveniences, and AI dissolves them without asking.

States that restrain fund less of everything, including compensation

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

Even where state-level restraint is politically achievable, the coalitions behind it weaken as the costs become visible. Capital moves, skilled people leave, competitive position erodes, all within a few years. The state either abandons restraint or accepts persistent relative decline.

Governments are competing to host the infrastructure rather than to slow it. OpenAI’s programme has expanded to the United Kingdom and, with Samsung and SK, to Korea. National security sharpens the pressure. AI is increasingly treated as military, intelligence, scientific, industrial and cyber capacity. A state that falls behind in AI falls behind in all of those, and no major state will accept that, whatever the social cost of deployment.

Friction changes the timing, not the equilibrium

The usual objection is that real deployment is slow. Firms face integration costs, liability, regulatory delay, internal politics, unions, sceptical customers, professional gatekeeping, data-protection duties and infrastructure limits. The objection accepts the competitive logic but argues that friction stops it operating at the speed the thesis implies.

Friction is real, and the thesis does not deny it. It denies that friction is a stable cooperative equilibrium. A firm can delay for reasons of liability, culture, integration cost or regulatory uncertainty. If unit cost dominance holds and its competitors deploy, the delay becomes a cost disadvantage. The firm still has to deploy in the end, or accept a slow loss of position. Friction moves the date of deployment. It does not decide whether deployment happens.

The same holds at every layer. Workers who delay are outcompeted by workers who do not. Sectors that delay are eroded by adjacent and foreign entrants. States that delay accept relative decline. In each case friction is a transient cost, not a permanent settlement. Friction as rescue would need to be persistent and universal, and it is neither, because the actors bearing the friction are competing with actors who are not. The final essay, Drag Is Not Rescue, develops this in full.

Coordination fails on observation and enforcement

The textbook answer to a Prisoner’s Dilemma is coordination. If players can communicate, commit and enforce, they can reach the outcome both prefer. Treaties, industry standards and binding agreements are the institutional forms. The question is whether any of them can hold restraint in place here.

The interaction is repeated, which helps. But the classical escape from the dilemma needs more than repetition. It needs defection to be observable and enforcement to be credible. Both are missing.

Defection is not observable because deployment leaves no outward trace. The product is the same, the output is the same, the customer interaction is the same. Seeing that the production process has changed means inspecting internal workflows, which firms have every reason to obscure and regulators lack the capacity to examine at scale. Payroll data, sector productivity statistics and wage patterns can reveal deployment afterwards. By then the defector has taken the advantage. Observation at the macro level is always after the fact, and enforcement that only works after the fact is not enforcement.

Enforcement is not credible because the actors able to enforce are the actors with the strongest reason to defect. States enforce international agreements. States are also the actors competing for AI advantage. A state that genuinely enforces restraint on its own firms hands ground to states that do not. The duty to enforce and the incentive to defect sit inside the same institution.

None of this makes coordination logically impossible. It makes coordination something other than an equilibrium, under weak enforcement, high strategic uncertainty and continuous technical change. Those are not passing features of the moment. They are structural features of how AI is being built and deployed.

The precedents worked under conditions that are absent here

Advocates of restraint point to earlier successes: nuclear non-proliferation, chemical weapons treaties, the phase-out of ozone-depleting substances. The successes are real. The conditions that produced them do not hold for AI.

Non-proliferation works as well as it does because the route to a weapon runs through scarce fissile material and identifiable facilities. AI has chokepoints too, in advanced chips, hyperscale data centres, cloud providers and energy. They matter for frontier training. They matter much less for the wage-demand problem, because labour substitution spreads through APIs, open-weight models, small specialised systems, enterprise wrappers and ordinary workflow tools. Slowing frontier training does not stop cognitive substitution below the frontier.

Chemical weapons treaties work because chemical weapons have almost no civilian economic value. A state that gives them up loses nothing in its economy. AI is the most economically valuable technology of the period, and a state that gives it up loses a great deal.

The ozone regime worked because the damaging chemicals had close substitutes that did the same job. Industry could switch without losing function. AI has no substitute for the cognitive work it performs except human cognition, which is exactly what it replaces because it is cheaper. The Montreal Protocol assumed an exit that does not exist here.

Nor is the race a Stag Hunt, because there is no shared stag. Defection by open-weight developers, foreign states, new entrants or individual workers destroys the cooperative outcome for everyone else. And it is not an arms race in the ordinary sense, because AI is not only a weapon. It is a productive asset. A state that restrains nuclear weapons keeps most of its economy. A state that restrains AI restrains its own growth, fiscal capacity and strategic infrastructure.

The precedents do not show that AI restraint is possible. They show that restraint has worked under specific conditions, and those conditions are absent.

The prize has no finish line

An ordinary Dollar Auction has a fixed prize. The game ends when someone wins the dollar.

The AI race does not end that way, because the prize is recursive. Each capability level unlocks the next. Each deployment produces revenue, feedback, workflow data, telemetry, distribution and political leverage that feed the next round. Winning one round changes the baseline for the next contest rather than ending the contest.

That makes exit impossible to calculate. In a finite game, a firm can model the endpoint and judge whether the remaining prize justifies the remaining spend. Here there is no final prize to calculate against. An actor that exits does not keep its position. It falls behind a frontier that keeps moving.

Nobody wins the race once. They survive a round, and the next round starts at once with higher stakes and shorter timelines. The absence of a finish line does not release the players. It removes the one thing that could have stopped them.

What would refute this

Premise Three fails if someone identifies an actor or coalition that can credibly restrain the propagation, and specifies how that restraint becomes a stable equilibrium rather than a temporary delay. The mechanism has to work at all four levels at once. It has to show how defection can be observed and punished when it is internal, diffuse, continuous and highly rewarded, and when the enforcers are also the actors most exposed to the cost of restraint.

Pointing to past treaties does not count unless it explains why the conditions that made them work hold for AI. Pointing to friction does not count, because friction modulates timing. The evidence that would count is a sustained, observable slowdown in deployment that holds while competitors could defect and do not, together with stable entry-level hiring in the sectors that restrained.

From restraint to definition

There is one more reason coordination fails, and it needs its own treatment. Even if observation and enforcement could be solved, a rule restraining AI would have to say what it restrains. That means drawing a line between AI that assists and AI that replaces. The next four essays make the case for Premise Four, that no available route preserves the wage-demand circuit. It begins with regulation, and with the discovery that inside a workflow the line does not exist.