The Treaty Paradox: Regulating Frontier Technologies Under Sovereign Disobedience

In June 1946, Bernard Baruch stood before the newly formed United Nations Atomic Energy Commission and presented Washington’s ambitious proposal for the international control of nuclear energy. The Baruch Plan proposed that all atomic activities—from uranium mining to reactor construction—be placed under the absolute monopoly of an international authority empowered to conduct unannounced on-site inspections without veto interference. "We are here to make a choice between the quick and the dead," Baruch famously declared.

Within two years, the plan collapsed into the Cold War’s first great diplomatic graveyard. The Soviet Union refused to permit foreign inspectors across its sovereign industrial perimeters, while the United States refused to dismantle its nuclear arsenal before verification was absolute.

Eighty years later, the international diplomatic community has convened once again in Geneva, London, and Seoul, attempting to resurrect Baruch’s ghost for a far more elusive technology: frontier artificial intelligence and synthetic biology.

From the Bletchley Park Declaration to the European Union’s push for an International AI Governance Agency modeled on the International Atomic Energy Agency (IAEA), international lawyers and diplomats operate under the enduring liberal illusion that catastrophic systemic risks can be domesticated through multilateral treaties.

Yet this diplomatic ambition rests upon a fatal category error. What international policy circles describe as the "global governance of frontier technology" is confronting what I term The Treaty Paradox: the geopolitical reality that the very structural properties that make frontier technologies existentially powerful render them fundamentally unverifiable under sovereign statecraft.

When the strategic asset is not a twenty-ton uranium centrifuge, but an ensemble of weight tensors that can be compressed onto a pocket-sized solid-state drive and trained inside air-gapped subterranean bunkers, traditional treaty mechanisms do not produce international security. They produce an asymmetric handicap—penalizing democratic nations that respect the rule of law while handing structural dominance to regimes practicing calculated sovereign disobedience.

The Collapse of the Physical Verification Trilemma

Every successful disarmament and non-proliferation treaty in modern history—from the 1968 Nuclear Non-Proliferation Treaty (NPT) to the 1993 Chemical Weapons Convention—rested upon an unshakeable geopolitical foundation: physical material asymmetry.

Nuclear weapons require rare, fissile isotopes (uranium-235 or plutonium-239) whose extraction demands massive industrial facilities: gaseous diffusion plants covering hundreds of acres, cooling towers emitting distinct thermal signatures, and gas centrifuge cascades with unmistakable electromagnetic and vibrational profiles.

Because physics cannot be encrypted, compliance can be verified through satellite remote sensing, environmental air sampling, and on-site seal placement.

The NPT worked not because states trusted one another, but because cheating was industrially conspicuous.

Frontier computational intelligence obliterates every premise of this verification paradigm. State sovereignty over frontier software breaks down across three insurmountable structural axes:

1. The Dual-Use Asymmetry

In nuclear physics, there is a distinct boundary between low-enriched reactor fuel (3–5% U-235) and weapons-grade material (>90% U-235). In frontier computation, no such physical threshold exists.

A high-density liquid-cooled datacenter housing 50,000 accelerated processing units (GPUs or TPUs) can train a state-of-the-art biological synthesis model on Monday, optimize a regional civilian power grid on Tuesday, and train autonomous cyber-warfare swarms on Wednesday. The hardware is indistinguishable; only the loss function and the training data change.

To verify that an advanced cluster is not executing restricted training runs, foreign inspectors would require continuous, real-time access to memory bus telemetry, model checkpoints, and proprietary datasets. No sovereign state possessing advanced intelligence capabilities will ever permit an international delegation to inspect the runtime state of its national computing infrastructure.

2. The Weight Mobility Vector

Once trained, the ultimate weapon of the computational era is entirely weightless. A frontier model capable of autonomously discovering novel bioreactor pathogens or executing autonomous offensive operations against electrical grid SCADA systems is not an immovable physical asset. It is a file of floating-point numbers.

It can be compressed, encrypted using post-quantum lattice cryptography, partitioned across decentralized peer-to-peer shards, or exfiltrated via optical laser links to unmonitored offshore sanctuaries.

A treaty cannot seal a border against weight vectors. If an inspector arrives at an enterprise campus in Shenzhen, Palo Alto, or Tel Aviv, how does the inspector certify that an unaligned checkpoint was not mirrored to a private air-gapped facility thirty seconds prior to their arrival?

3. The Inference Evaporator

Even if training compute could be globally monitored through chip-level telemetry—a dystopian technical proposition fraught with supply-chain vulnerabilities—inference execution can be performed on consumer-grade hardware through 4-bit quantization, sparse mixture-of-experts routing, and distributed local meshes.

A prohibited sovereign model does not require a nuclear reactor to execute its inference loops; it requires only a few kilowatts of electricity and a bank of local accelerators operating off-grid.

The Game Theory of Calculated Disobedience

When verification is impossible, treaty dynamics degenerate into a classical Prisoner’s Dilemma with negative-sum payoffs for compliance.

In international relations theory, treaties function as commitment devices only when the expected cost of detection and sanction exceeds the sovereign advantage of defection. In frontier technologies, this calculus is inverted.

If State A (a democratic federation bound by parliamentary oversight, judicial review, and investigative journalism) signs a universal treaty banning frontier recursive self-improvement models or autonomous lethal coordination, State A must enforce the ban through domestic criminal statutes, public regulatory disclosures, and civil liability. Its domestic research labs, capital markets, and national security agencies are legally constrained.

Conversely, State B (an authoritarian competitor operating with a closed domestic media ecosystem and an integrated military-civil fusion apparatus) signs the identical convention with solemn diplomatic fanfare.

State B calculates—correctly—that the international community possesses zero forensic capability to detect clandestine training runs conducted inside secure underground facilities.

Furthermore, State B knows that if an unaccredited NGO or foreign intelligence service alleges a violation, State B can simply dismiss the accusations as politically motivated espionage, invoke sovereign defense prerogatives, and block inspection teams under standard national security exception clauses.

"A treaty that cannot be verified is not a peace pact; it is a mechanism for unilateral disarmament. It imposes legal paralysis upon those who respect the law, while offering absolute impunity to those who disregard it."

The payoff matrix is merciless. The compliant state incurs massive economic and technological handicaps, falling behind in the compute frontier and losing defensive superiority. The non-compliant state reaps the compounding dividends of unrestricted frontier capabilities while enjoying the moral camouflage of treaty signatory status. Under such conditions, sovereign disobedience is not an aberration; it is the dominant rational strategy.

The Illusion of Zero-Knowledge Sovereign Audits

In response to the verification crisis, technocratic advocates of "verifiable multilateralism" have proposed cryptographic panaceas, most notably Zero-Knowledge Proofs (ZKPs) and hardware-enforced Trusted Execution Environments (TEEs).

The argument, championed by think-tank technologists in Brussels and Washington, suggests that sovereign states could install cryptographic coprocessors on all cutting-edge silicon wafers at the foundry stage. These coprocessors would cryptographically sign training logs and generate zero-knowledge mathematical proofs certifying that a cluster’s aggregate FLOP count never exceeded regulatory caps (e.g., $10^{26}$ operations) and that prohibited loss functions were never executed—all without revealing the underlying data or model architecture.

As an abstract mathematical theorem, zero-knowledge verification is elegant. As a mechanism of sovereign statecraft, it is laughably naïve.

First, hardware-level verification requires trusting the foundry root of trust. Who holds the cryptographic signing keys for the silicon governors? If the keys are held by an international body, every major power will immediately treat that body as an espionage conduit for foreign adversaries. If the keys are held domestically, the sovereign state possesses infinite incentive and capability to execute physical side-channel attacks, microcode patches, or clock-manipulation exploits to bypass the hardware counters.

Second, the supply chain for advanced semiconductors is already a kinetic geopolitical battleground. No major power will permit its national security compute infrastructure to be beholden to hardware kill-switches or remote cryptographic attestations manufactured by foreign foundries.

Third, zero-knowledge proofs verify only that a specific mathematical constraint was satisfied within the hardware perimeter. They cannot verify what happens outside the chip. They cannot detect whether training datasets were contaminated with synthetic bioweapons blueprints, nor can they prevent a state from distributing a forbidden training workload across two hundred thousand un-metered secondary chips operating below the regulatory detection threshold.

The belief that cryptography can substitute for geopolitical trust is the final retreat of multilateralism when confronted with sovereign intractability.

From Universal Treaties to Minilateral Export Cartels

If universal, comprehensive non-proliferation treaties for frontier technologies are structurally dead on arrival, what replaces them?

As I demonstrated in The Sunset of Consensus, the era of universal 193-state United Nations conventions has run its course. In its place, sovereign governance over transformative technologies is coalescing around minilateral export cartels and industrial alliances.

Rather than seeking unattainable consensus at the UN General Assembly, like-minded technological powers are constructing closed, fortified perimeters based on tangible physical chokepoints:

1. Foundry and Lithography Chokepoints

While software and model weights are infinitely portable, the industrial machinery required to manufacture leading-edge silicon is brutally concentrated. Extreme Ultraviolet (EUV) photolithography systems rely on thousands of optical and laser components manufactured by a handful of companies across the Netherlands, Japan, Germany, and the United States.

Governance will not be enforced by UN treaties; it will be enforced through minilateral export control clubs (such as the Chip 4 Alliance and modernized COCOM regimes) that weaponize monopoly control over physical lithography to deny hostile regimes the hardware foundations of frontier intelligence.

2. Verified Sovereign Enclaves

Allies will establish mutual zero-trust verification zones. Instead of asking adversaries to verify compliance, allied states will pool computational resources, intelligence feeds, and defensive countermeasures within an integrated sovereign umbrella.

Within this perimeter, mutual inspection is enforceable through reciprocal economic interdependence, joint military command structures, and shared intelligence agreements (such as the Five Eyes). Outside the perimeter, sovereign disobedience is assumed as a default operational fact.

3. Defensive Deterrence Over Regulatory Bans

The international system must recognize that offensive frontier capabilities—whether automated vulnerability discovery or autonomous drone swarm coordination—cannot be regulated out of existence.

The sole effective deterrent against an unconstrained sovereign actor deploying a frontier autonomous weapon is not the fear of an International Court of Justice condemnation; it is the certainty that the targeted coalition possesses automated, self-healing cyber and physical counter-systems operating at equal or superior speed.

The Courage of Skepticism

The siren song of universal treaties is seductive because it promises safety without sacrifice. It allows politicians to sign eloquent declarations in European palaces, secure laudatory headlines, and reassure their electorates that the frontier has been tamed.

It is a dangerous delusion.

By placing our faith in uninspected paper conventions, we repeat the catastrophic complacency of the 1928 Kellogg-Briand Pact, which solemnly outlawed war on parchment while the totalitarian machines of the twentieth century prepared for global conquest.

Frontier technology is a test of civilizational realism. The states that will survive and protect human freedom in the coming decades are not those that draft the most comprehensive regulatory manifestos, but those that possess the sober courage to acknowledge the limits of diplomacy: to recognize that when an existential capability cannot be verified, peace is maintained not by treaties signed in bad faith, but by sovereign vigilance, material industrial resilience, and the relentless defense of democratic ground truth.