Redistributed

The Cold War produced the most consequential piece of game theory in human history: the doctrine of mutually assured destruction. Two superpowers, each capable of annihilating the other, reached a stable equilibrium in which neither used its weapons — precisely because the weapons existed. Deterrence worked not through goodwill but through mathematics. The logic was brutal and elegant: any first strike would trigger a retaliatory second strike, so no rational actor would strike first.

Now a growing number of strategists ask whether this logic can be transplanted to artificial intelligence. Renaud Bougueng T., in his 2026 strategic analysis AI and Human Civilization, proposes "artificial dissuasion" — which he defines as "a game-theoretic maturity model to prevent lose-lose autonomous deployments, inspired by but distinct from nuclear deterrence." The concept anchors restraint not in treaties alone but in "cryptographic mechanisms... to immutable universal properties," creating verifiable commitments that make reckless deployment costly for everyone, including the deployer.

It is a serious proposal for a serious problem — and Bougueng deserves credit for attempting to build a rigorous analytical framework rather than relying on analogies alone. The proposal also runs into difficulties that reveal as much about AI governance as they do about deterrence theory.

How Nuclear Deterrence Actually Worked

The standard account of nuclear deterrence is deceptively simple, so it is worth specifying the conditions that made it function.

Mutually assured destruction (MAD) required second-strike capability — the guarantee that even after absorbing a full nuclear attack, a nation could still launch a devastating response. The United States achieved this through a triad of land-based missiles, submarine-launched missiles, and strategic bombers. The Soviet Union matched it. Destroying all three legs simultaneously was impossible, so retaliation was assured.

Rational actors with identifiable addresses. The logic assumed that decision-makers valued their own survival. It also assumed you knew who the decision-makers were. Nuclear weapons required state-scale industrial capacity — enrichment facilities, delivery systems, testing infrastructure. There were, at the peak, nine nuclear-armed states. Attribution was straightforward: if a missile launched from Soviet territory, Moscow was responsible.

Clear escalation ladders. Herman Kahn's famous (and deliberately provocative) "escalation ladder" described 44 rungs from subcrisis disagreement to total nuclear war. This granularity mattered because it gave adversaries room to signal, de-escalate, and negotiate at each step. The 1962 Cuban Missile Crisis resolved precisely because both sides found off-ramps that preserved deterrence without triggering war.

Arms control as stabilizer. The system was not pure brinkmanship. A dense network of treaties — the Limited Test Ban Treaty (1963), the Non-Proliferation Treaty (1968), SALT I and II, START — created verification mechanisms, imposed limits, and established communication channels. The Moscow-Washington hotline, installed after the Cuban crisis, was infrastructure for preventing miscalculation.

This architecture took decades to build. It survived several near-misses — Stanislav Petrov's 1983 decision to override a false alarm, the Able Archer 83 exercise that nearly triggered Soviet preemption. Deterrence "worked" in the sense that nuclear weapons were not used after 1945. Whether this was because of the doctrine or despite it remains genuinely debated among strategists.

Where the Analogy Holds

Bougueng's artificial dissuasion framework draws on real structural parallels between nuclear weapons and autonomous AI systems.

Catastrophic, irreversible consequences. Both nuclear weapons and certain AI deployments involve decisions that cannot be undone. A nuclear launch is irrevocable. An autonomous AI system deployed into critical infrastructure — financial markets, military command-and-control, power grid management — could trigger cascading failures that propagate faster than human intervention can contain. The shared feature is that the cost of getting it wrong is not incremental but catastrophic.

The value of credible commitments. Nuclear arms control worked because commitments were verifiable. On-site inspections, satellite monitoring, seismic detection networks — each created mechanisms for trust that did not depend on trust. Bougueng's proposal for cryptographic verification mechanisms serves the same function: creating machine-verifiable constraints on AI deployment that do not rely on good faith alone. If an AI system's operational parameters can be cryptographically attested, adversaries can verify compliance without revealing proprietary details.

Game-theoretic structure. Both domains feature coordination problems where unilateral action is tempting but collectively destructive. The prisoner's dilemma in AI safety investment is well-documented: any single company that slows down to invest in safety risks losing market position to competitors who do not. Deterrence reframes this dynamic — if reckless deployment triggers costly retaliation (regulatory, economic, or technical), the incentive to cut corners diminishes.

International coordination as force multiplier. Confidence-building measures, transparency mechanisms, and shared norms reduced the risk of nuclear miscalculation. The same logic applies to AI: international AI safety summits, whatever their limitations, create channels for communication that reduce the risk of competitive dynamics spiraling into destructive deployment races.

Where the Analogy Breaks

The parallels are real. The differences are more consequential.

Attribution is the foundational problem. Nuclear weapons have return addresses. If a warhead detonates, seismic data, satellite imagery, and fallout analysis identify the source within hours. AI deployments are diffuse, dual-use, and often invisible. When an algorithmic trading system triggers a flash crash, or an autonomous hiring system discriminates at scale, identifying the responsible actor — let alone attributing intent — is genuinely difficult. Deterrence without attribution is a gun without a trigger.

The actor problem. Nuclear deterrence operated among a small number of nation-states. AI development involves states, corporations, research laboratories, open-source communities, and individual developers. OpenAI, Google DeepMind, Anthropic, Meta, Mistral, dozens of Chinese labs, thousands of fine-tuners, millions of deployers. Deterrence theory assumes you can identify, communicate with, and credibly threaten the relevant decision-makers. In AI, it is not always clear who they are.

The cost barrier has inverted. Building a nuclear weapon required massive state investment — the Manhattan Project cost roughly $30 billion in today's dollars. This natural barrier to proliferation made the nuclear club small and manageable. AI capabilities are moving in the opposite direction. Export controls on advanced chips attempt to reinstate a cost barrier, but the trend toward more efficient training, open-weight models, and algorithmic improvements means that meaningful AI capability is becoming more accessible, not less. Deterrence that depends on scarcity confronts a technology defined by abundance.

No second strike. The core mechanism of nuclear deterrence — guaranteed retaliation — has no clear AI equivalent. If a nation deploys a reckless autonomous weapons system, what is the deterrent response? Deploying your own reckless system? That is not deterrence; it is an arms race. Economic sanctions? Those operate on timescales of months, not milliseconds. The absence of a credible, automatic retaliatory mechanism undermines the game-theoretic foundation that made nuclear deterrence stable.

Continuous, not binary. Nuclear weapons present a binary choice: use or don't use. This clarity is what makes deterrence possible — the threshold is unambiguous. AI deployment is continuous and incremental. There is no bright line between a helpful AI assistant, an autonomous agent, and a system capable of causing catastrophic harm. The escalation ladder has no clearly marked rungs. When does a competitive AI deployment cross from aggressive to reckless? Who decides?

Time to stability. Nuclear deterrence took roughly fifty years to reach a precarious equilibrium — and that equilibrium included the Cuban Missile Crisis, Able Archer 83, and multiple incidents where individual judgment prevented catastrophe. AI governance does not have fifty years. The technology is advancing too quickly and the actors are too numerous for a half-century shakeout period.

What Game Theory Actually Suggests

If deterrence maps poorly onto AI, other game-theoretic frameworks map better.

Coordination games, not deterrence. The fundamental challenge in AI governance is not threatening adversaries into compliance. It is coordinating among actors who would all benefit from restraint but cannot achieve it unilaterally. This is a coordination game, not a deterrence problem. The solution is not mutual threat but mutual commitment — shared standards, common evaluation frameworks, interoperable safety protocols.

Ostrom over Schelling. Elinor Ostrom's work on governing the commons offers a more productive framework than Thomas Schelling's deterrence theory. Ostrom demonstrated that common-pool resources — fisheries, forests, irrigation systems — can be governed effectively without either privatization or state coercion, provided certain institutional conditions are met: clear boundaries, proportional costs and benefits, collective decision-making, monitoring, graduated sanctions, conflict resolution mechanisms (Ostrom, 1990). AI safety has the structure of a commons problem. The "commons" is the shared information environment, the labor market, the stability of critical infrastructure — resources degraded by reckless AI deployment and preserved by responsible governance.

Race-to-the-bottom dynamics. The prisoner's dilemma in AI safety spending is well-established. As Daron Acemoglu has argued, firms can pursue "so-so automation" — systems that replace workers without meaningful productivity gains — because the costs of displacement are externalized onto workers and public safety nets while the savings accrue to shareholders (Acemoglu, 2021). Without coordination mechanisms, competitive pressure drives toward faster deployment with less safety investment. This is not a deterrence failure. It is a market failure that requires regulatory solutions: mandatory safety evaluations, liability frameworks, clear regulatory authority.

The Redistribution Problem with Deterrence Thinking

Here is the part that deterrence frameworks tend to obscure: deterrence entrenches existing power holders.

The Non-Proliferation Treaty is the canonical example. Signed in 1968, it created two classes of states: those permitted to possess nuclear weapons (the five that already had them) and those forbidden from acquiring them. The nuclear powers wrote the rules, and the rules preserved their monopoly. They promised disarmament in Article VI. Fifty-eight years later, all five original nuclear states maintain and modernize their arsenals.

Apply this logic to AI. A deterrence framework for artificial intelligence — particularly one that emphasizes verification, compliance mechanisms, and credible threats — will be designed by and for the actors who currently dominate AI development. That means a handful of US and Chinese companies, and the governments behind them. Nations pursuing sovereign AI capacity would be constrained. Open-source communities would face compliance burdens. Researchers in the Global South would navigate frameworks designed in Washington and Beijing.

Bougueng is aware of this risk. His framework emphasizes "universal properties" and proposes cryptographic mechanisms that are, in principle, neutral — verifiable by any party. But the history of arms control suggests that verification regimes are never neutral in practice. The IAEA's inspection capabilities favor states with the diplomatic leverage to shape inspection protocols. Export control regimes like the Wassenaar Arrangement reflect the strategic priorities of their founding members. China's exclusion from Western AI governance frameworks is not an accident — it is a feature of frameworks designed by China's strategic competitors.

The deeper issue is that deterrence thinking frames AI governance as a problem of managing threats between powerful actors. But the most significant redistribution questions in AI are not about preventing catastrophic deployment by rival superpowers. They are about who benefits from ordinary deployment — whose jobs are displaced, whose data is extracted, whose communities bear the costs of algorithmic decision-making. Institutional decoherence is not caused by adversarial intent. It is caused by the mundane inability of governance to keep pace with deployment.

A framework borrowed from nuclear strategy may address the dramatic tail risk — autonomous weapons, catastrophic AI failures — while leaving the quotidian redistribution problems untouched. And it is the quotidian problems that affect the most people.

What This Means for You

If you work in AI governance: Bougueng's artificial dissuasion framework names a real problem — the need for credible, verifiable commitments to prevent destructive AI deployment. Take the game-theoretic rigor seriously. But do not let deterrence metaphors crowd out the governance work that matters most: liability rules, transparency requirements, labor protections, and democratic oversight mechanisms. Ostrom beats Schelling for most AI governance problems.

If you work in policy: be wary of frameworks that replicate nuclear governance's power dynamics. The NPT preserved a five-member nuclear club. AI governance frameworks should expand the circle of meaningful participation, not narrow it. Strategic foresight documents are useful starting points, but the institutional design matters more than the strategic vision.

If you are a citizen: the debates about AI deterrence and safety summits are important, but they should not distract from the immediate governance questions that shape your life — whether your employer can use AI to set your wages, whether your insurer can use AI to price your premiums, whether your government can use AI to allocate public benefits. Those are redistribution questions. They do not require deterrence theory. They require democratic politics.

Sources

Acemoglu, Daron. "Harms of AI." Cambridge Journal of Economics, 2021. Cambridge
Bougueng T., Renaud. AI and Human Civilization: Strategic Foresight. 2026.
Kahn, Herman. On Escalation: Metaphors and Scenarios. Praeger, 1965.
Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990. Cambridge
Schelling, Thomas C. The Strategy of Conflict. Harvard University Press, 1960. Harvard
Winner, Langdon. "Do Artifacts Have Politics?" Daedalus 109, no. 1 (1980): 121–136. JSTOR

General

Artificial Dissuasion: Can Nuclear Deterrence Logic Apply to AI?