Can humanity pause artificial intelligence?
The question now comes from technology executives, security officials, and religious leaders alike. Anthropic has argued that as frontier systems cross defined capability thresholds, labs and governments may need pre-agreed mechanisms to slow or halt development until safeguards catch up. Pope Leo XIV, in his first encyclical, Magnifica Humanitas, has warned that AI must not be allowed to erode human dignity, above all where it touches war and the fate of human life. These warnings deserve to be taken seriously. AI is moving into finance, healthcare, intelligence, public administration, cyber operations, and military planning. It can write, code, analyse, persuade, deceive, predict, and increasingly act. The fear that it may outpace the institutions meant to govern it is legitimate.
The Pause Problem
But the pause debate rests on a fragile assumption: that the world still has the political capacity to stop.
The strongest version of the argument is worth engaging, because the weak one is easy to dismiss. Few credible voices want to freeze all AI everywhere and forever. The thoughtful proposal is narrower: a conditional, coordinated slowdown at the frontier, triggered by agreed risk thresholds, applied to the most capable systems, verified among the handful of actors who can build them. It is this version, not the cartoon of a global off switch, that geopolitics is quietly dismantling.
The flaw behind the pause argument is not moral but political. A conditional pause assumes states will restrain a technology they increasingly associate with survival, influence, and military advantage. Yet the race is driven by states that fear dependency, militaries that fear surprise, and governments that fear strategic inferiority. Restraint then becomes a question of trust rather than engineering, and trust is precisely what is missing. Would Washington accept a freeze it could not confirm Beijing was honouring it? Would Beijing pause while American firms held a lead? Each actor must slow down today against the unverifiable possibility that a rival already has not.
History offers a sobering template. Nuclear weapons were pursued less from scientific ambition than from the fear that rivals would acquire them first, and every move each side made in the name of security deepened the other’s insecurity. AI risks the same pattern, with one decisive difference, and the difference cuts both ways. As a governance model, nuclear arms control is the right template: red lines, declared thresholds, verification where possible, crisis channels. But as an enforcement mechanism, the analogy collapses. A nuclear warhead programme needs rare materials and visible facilities; you can inspect a centrifuge cascade. AI needs data, algorithms, compute, and talent, all distributed across the global economy. You cannot inspect a model the way you inspect a reactor. The framework transfers; the leverage does not.
Shifting Governance
The evidence is already on the table. When Washington tightened controls on advanced chips to slow China’s progress, Beijing did not stop. Within a year, SMIC was producing 7-nanometre chips for Huawei’s Mate 60 Pro , hardware the controls were meant to block, and capability migrated rather than disappeared. It poured resources into domestic alternatives, and capability migrated rather than disappeared. In Ukraine, cheap drones and AI-assisted targeting reshaped the battlefield faster than any treaty could regulate. The same logic surfaced even when a halt was imposed deliberately: in June 2026, when the US Commerce Department forced Anthropic to suspend its most capable models, the company could only comply by disabling them for everyone, and noted that the flagged vulnerability was already reproducible in rival systems. Knowledge, once spread, cannot be uninvented. The race is powered by a distributed ecosystem of capital, fear, and expertise that no single regulator commands.
None of this makes governance useless. It shifts the task from stopping AI to controlling its most dangerous uses.
Here a fair objection arises, the same one that sinks the pause. If frontier capability cannot be verified, why would any other restraint hold? The answer is that the proposals worth defending do not depend on the trust a pause requires. A freeze asks each party to forgo advantage today on faith that rivals are doing the same; it fails the moment one suspects defection. The restraints below are different in kind, because each party benefits from observing them regardless of what the other does.
The first is meaningful human control over lethal decisions. AI already supports drone swarms, target identification, and battlefield logistics, and it compresses crisis decision-making from hours into minutes. A binding norm that a human must authorise any use of lethal force, with the time and information to do so, is the single red line most worth defending. No commander needs to trust an adversary to want it: a military that lets its own systems escalate without authorisation endangers itself first.
The second is crisis communication between major powers. Arms control survived the Cold War less through treaties than through channels that prevented adversaries from misreading each other in the moments that mattered. The AI equivalent is urgent: protocols for when an automated system triggers an alert, and a direct line to de-escalate before algorithmic speed forecloses human judgment. Like the lethal-force norm, this serves every party’s interest in not stumbling into a war neither chose. Because each restraint is self-enforcing, adopting it requires none of the trust a pause demands, and each instance of mutual observance builds the thin track record from which deeper cooperation can grow.
Drafting the Pause in the Middle
This is where middle powers find their role, and where Australia has more leverage than its place in the chip race suggests. Canberra will not build the frontier, but it sits inside the Indo-Pacific, the theatre where these norms are most likely to be tested first. Middle powers cannot enforce a pause, but they can convene, draft, and legitimise the standards that adversaries will not propose to one another directly. The incentive to adopt them comes not from the convener but from the standards themselves: because each restraint serves its adopter regardless of what rivals do, a middle power’s role is simply to supply the neutral language through which both sides can embrace limits they already have reason to want, without appearing to concede. Norm entrepreneurship of this kind is not a consolation prize for those outside the race; it is the part of the architecture the leaders are too distrustful to build alone.
So, the determinism is partial. Markets, militaries, and the spread of knowledge have largely settled whether the race continues. What remains open is how it is run, and that is the space governance can still occupy. The task is not to pretend history can be frozen, but to write the rules before rivalry writes them for us. Such rules will not stop the race. They may keep it from becoming reckless. The real question was never whether humanity would continue; fear and competition have made that choice. It is whether we can govern AI before we automate restraint out of existence.
Muhammad Amir is a PhD researcher at Deakin University, specializing in international relations and security studies. His research focuses on peace processes, strategic competition, defence policy, and emerging technologies.
This article is published under a Creative Commons license and may be republished with attribution.
