Can Artificial Intelligence Be Paused?

By Matthew Parish

Wednesday 9 September 2026

There is an increasingly fashionable proposition in discussions about artificial intelligence that humanity ought to stop for a while. Artificial intelligence is developing so quickly, the argument runs, that governments, societies and even the scientists constructing the systems themselves cannot understand all the consequences of what they are doing. Therefore we ought to pause the development of increasingly powerful systems, perhaps for six months, perhaps for a year or perhaps for longer, while we decide what rules ought to govern them. The argument has an immediate intuitive appeal. If humanity is constructing something extraordinarily powerful whose ultimate properties she does not entirely understand, surely the prudent course is to stop periodically and examine what has been built before proceeding further.

The difficulty is that technological history rarely works like this. Humanity does not generally invent something, collectively contemplate its consequences and then decide whether to continue. Technologies emerge competitively. Individuals, corporations, universities, intelligence agencies, armies and states pursue advantages over one another, while the rules governing new technologies usually appear afterwards. The steam engine was not paused while governments considered industrialisation. Nuclear physics did not stop while diplomats considered the implications of the atomic bomb. The internet did not suspend its development while legislators decided what social media might do to democratic politics. Artificial intelligence is following much the same pattern, except extraordinarily quickly. This creates a disturbing possibility: an AI pause might be highly desirable while at the same time being politically and technologically impossible.

That tension deserves serious consideration because the case for slowing down is far from foolish. Artificial intelligence is advancing at an extraordinary rate. Stanford’s 2026 AI Index concludes that capability is continuing to accelerate rather than plateau. Industry produced more than 90 per cent of notable frontier models in 2025, while systems reached or exceeded human baselines on increasingly demanding scientific, mathematical and reasoning tests. At the same time the report identifies a growing discrepancy between capability development and responsible-AI practices, while documented AI incidents increased substantially. The uncomfortable conclusion is that the machines are improving more quickly than the institutions intended to supervise them.

There are several reasons why this ought to concern us. The first is economic. Earlier technological revolutions predominantly mechanised physical labour or relatively repetitive intellectual work. Artificial intelligence is different because its natural territory is cognition itself. Lawyers, programmers, accountants, financial analysts, translators, doctors, journalists, consultants, designers and academics are discovering that increasingly sophisticated parts of their work can be undertaken by machines. There is no reason to assume that every one of these professions will disappear and historical experience suggests that technological revolutions create occupations as well as destroying them. Nevertheless there is also no economic law requiring the new occupations to employ the same people, in the same places, at the same salaries or with the same social status as the occupations that disappear. An extremely rapid transition might therefore generate enormous aggregate wealth while simultaneously producing profound individual insecurity and political dislocation.

The second concern is the concentration of power. Frontier artificial intelligence is extraordinarily expensive. It requires huge quantities of computing capacity, advanced semiconductors, electricity, specialist personnel, data-centre infrastructure and capital. Stanford estimates that global AI compute capacity has been growing approximately 3.3-fold annually since 2022, while the United States hosts more AI data centres than any other country by an enormous margin. The hardware supply chain is itself remarkably concentrated, with the majority of the most advanced AI chips dependent upon fabrication by TSMC in Taiwan. The intellectual infrastructure of twenty-first century civilisation might therefore come to depend upon a relatively small number of corporations operating extraordinarily expensive machines. There are legitimate questions about whether concentrations of private power on this scale are compatible with the political assumptions upon which liberal democracies have traditionally operated.

The third concern is more fundamental. Artificial intelligence is becoming agentic. The distinction between a machine that answers questions and a machine that undertakes tasks is enormously important. Once an artificial intelligence system can write software, use computers, communicate with other systems, operate financial accounts, conduct research, manipulate networks and pursue extended objectives, we are no longer discussing merely an extraordinarily sophisticated encyclopaedia. We are discussing artificial entities capable of acting upon the world. The fact that such systems remain unreliable does not necessarily make the problem less serious. Stanford describes the contemporary state of AI as a “jagged frontier”: systems can display extraordinary competence in sophisticated fields while continuing to make apparently elementary mistakes. Agents have improved dramatically on computer-use benchmarks but still fail a significant proportion of tasks. Powerful intelligence combined with intermittent incompetence may sometimes be more dangerous than consistent mediocrity.

Hence the strongest argument for a pause is not that artificial intelligence is inherently malevolent or that technological progress ought to be stopped. It is that there may be enormous value in allowing political institutions to catch up. A year without another dramatic increase in frontier capability could be used to develop testing regimes, liability rules, cybersecurity standards, international protocols, rules concerning autonomous weapons and institutions for independent evaluation. Governments might decide that certain decisions must always remain subject to meaningful human control, particularly decisions concerning nuclear weapons, lethal force, criminal punishment and other exercises of coercive state authority. Societies might also begin adapting educational systems and labour markets to a world in which intellectual scarcity becomes progressively less important.

The analogy frequently invoked is nuclear arms control. The United States and Soviet Union were mortal geopolitical adversaries and neither trusted the other. Nevertheless they eventually recognised that certain technological competitions generated dangers to both sides. Arms control did not depend upon friendship. On the contrary, it depended upon institutionalised suspicion: treaties, inspections, satellite surveillance, declarations, verification procedures and consequences for cheating. Something similar might theoretically be constructed for artificial intelligence. Governments could agree that systems exceeding specified computational thresholds would require registration, notification or inspection. Extremely large training runs might have to be declared. Advanced semiconductor shipments could be monitored. Data centres above specified sizes might become subject to regulatory obligations.

There is even a physical foundation upon which such an arrangement might be built. Artificial intelligence may ultimately consist of software but frontier AI cannot presently be developed using an ordinary laptop concealed in somebody’s bedroom. It depends upon immense computing clusters, specialised chips, electricity, cooling infrastructure and sophisticated semiconductor manufacturing. Those things leave footprints. This makes some degree of regulation possible. The semiconductor supply chain contains particularly conspicuous bottlenecks, while enormous data centres are difficult to conceal. A system of compute governance might therefore establish a temporary ceiling above which neither American nor Chinese companies could train new systems without mutual notification or agreed safeguards.

The geopolitical difficulty is nevertheless formidable because the United States and China are now remarkably close competitors. Stanford’s 2026 figures indicate that American institutions produced 59 notable models in 2025 compared with China’s 35, but China leads in AI publication volume, citations and patent grants, while the performance difference between leading American and Chinese models has narrowed dramatically. American and Chinese models have repeatedly exchanged positions at the technological frontier since early 2025. This is no longer a technological competition in which one side enjoys an overwhelming lead and can afford magnanimous restraint. It is a strategic race in which both sides have reasons to fear that a temporary pause might permanently disadvantage them.

The logic is essentially a prisoner’s dilemma. Suppose Washington required American laboratories to cease training substantially more powerful models for twelve months. Beijing might make reassuring noises about responsible innovation while Chinese laboratories continued working. At the end of the year China might possess a decisive technological advantage in a field relevant not merely to commercial productivity but to cyber warfare, intelligence analysis, autonomous weapons, military logistics, scientific research and perhaps strategic planning itself. No responsible American administration could comfortably accept that risk. Yet exactly the same argument applies in reverse. If Beijing ordered Chinese laboratories to pause while Silicon Valley continued accelerating, the Chinese government would fear that it had voluntarily surrendered one of the most consequential technological contests of the century.

There are therefore respectable arguments not merely that a pause is impractical but that it might actually be dangerous. The development of artificial intelligence in the United States takes place, however imperfectly, within a political system containing courts, an independent press, competing corporations, universities, civil society organisations and democratic institutions. China possesses a radically different political structure. If Western democracies imposed severe restraints upon themselves while authoritarian states did not, the consequence might not be a safer world. It might instead be a world in which the most powerful artificial intelligence systems were disproportionately developed by governments less constrained by liberal ideas concerning privacy, political freedom and individual autonomy. An unsuccessful pause could therefore achieve precisely the opposite of its intended objective.

There is also a more elementary economic problem. Artificial intelligence may generate extraordinary increases in productivity and scientific knowledge. A pause would therefore have costs measured not merely in corporate profits but potentially in discoveries delayed. AI systems are increasingly being used in medicine, chemistry, materials science, engineering and other fields in which accelerated research might save lives. If artificial intelligence eventually contributes substantially to new medicines, better climate technologies, improved agricultural productivity or solutions to currently intractable scientific problems, deliberately slowing its development carries a moral cost of its own. We should not assume that risk exists only on the side of technological acceleration. There are also risks associated with technological stagnation.

Competition within countries creates another difficulty. Even if Washington and Beijing reached an agreement, they would somehow have to persuade their own companies, researchers and military establishments to comply. The incentives to cheat could be enormous. A corporation believing that its competitors had stopped might see an unprecedented opportunity to obtain technological supremacy by continuing secretly. Intelligence agencies would have obvious reasons to pursue classified research. Military establishments might insist upon exemptions on national security grounds. Universities might argue that restrictions upon research violated academic freedom. Startups would search for loopholes. Capital would move between jurisdictions. The larger the commercial and strategic prize, the greater the incentive to circumvent whatever definition of a “pause” governments adopted.

The problem becomes still more serious as artificial intelligence becomes cheaper. Today’s frontier model requires extraordinary computing infrastructure but yesterday’s frontier model does not. Algorithms improve, chips become more efficient, knowledge disseminates and techniques discovered inside expensive laboratories eventually become public. Open-source and open-weight AI development continues to expand rapidly and millions of AI-related software projects already exist. A regulatory regime based entirely upon controlling giant computing clusters might therefore work for a time, but the threshold of capability obtainable using smaller clusters will continue rising. The physical bottleneck upon which enforcement depends may progressively weaken.

Nor are the United States and China actually the only actors. They dominate frontier model development but artificial intelligence research is increasingly international. Europe, Japan, South Korea, India, the Gulf states and numerous other countries possess significant scientific capabilities and substantial incentives to develop domestic AI industries. Open-source development further distributes knowledge internationally. An American-Chinese agreement might therefore initially restrain the two principal competitors only to create opportunities for a third country. The more successful artificial intelligence becomes, the more governments will regard sovereign AI capability as analogous to energy security, telecommunications infrastructure or military independence. A permanent global cartel suppressing frontier development would become progressively harder to maintain.

These practical objections distinguish artificial intelligence from nuclear weapons in a fundamental way. Constructing a nuclear weapon requires rare materials, specialised industrial facilities and conspicuous physical processes. The knowledge necessary to create increasingly sophisticated artificial intelligence is ultimately information. Information leaks, migrates, reproduces and becomes cheaper to implement. A government can confiscate enriched uranium. It cannot easily confiscate an algorithm once thousands of people understand it. Nuclear arms control may therefore be an illuminating analogy for AI governance, but it is also misleading if taken too literally. Artificial intelligence possesses characteristics of both nuclear technology and mathematics, and mathematics has never proved susceptible to international prohibition.

The contemporary relationship between Washington and Beijing makes the problem still more acute. The two countries are discussing AI safety even as their broader technological rivalry intensifies. Bilateral AI safety talks are expected to address matters including AI-enabled cyber threats and mechanisms for cooperation between laboratories. Yet almost simultaneously American authorities have accused Chinese AI companies of using American systems to accelerate their own technological development through large-scale distillation, allegations Beijing denies. Whatever the merits of those particular accusations, the episode illustrates the problem perfectly. The countries that would have to trust one another sufficiently to pause are simultaneously engaged in a technological competition characterised by allegations of copying, export restrictions, strategic rivalry and mutual suspicion.

This does not mean that international AI diplomacy is pointless. On the contrary, it suggests that policymakers should distinguish between the attractive but probably impossible objective of stopping artificial intelligence and the more realistic objective of slowing or prohibiting particular dangerous applications. The United States and China might conceivably agree that nuclear command-and-control systems should never delegate final launch authority to artificial intelligence. They might establish communications protocols for major AI incidents, just as nuclear powers developed hotlines. They might exchange information about systems that unexpectedly escape controlled environments or conduct autonomous cyber operations. Governments might establish common standards for testing models capable of assisting biological weapons development or sophisticated cyber attacks. These are narrower objectives and therefore considerably more plausible.

A temporary moratorium upon a very specific category of system might also occasionally be achievable. If credible evidence emerged that a particular capability created an immediate catastrophic risk, governments might agree not to deploy it while safeguards were constructed. The history of international relations contains examples of rivals accepting narrowly defined restraints when their interests genuinely converge. But this is different from the sweeping proposition that humanity might simply stop developing more powerful artificial intelligence until everyone feels comfortable proceeding again. There will never be universal agreement about when that moment has arrived.

This reveals the deeper problem with the idea of an AI pause. It assumes that humanity is a political actor. She is not. Humanity does not possess a government, a legislature or a police force. She consists of approximately eight billion people organised into competing states, corporations, universities, armies, intelligence agencies and communities, all possessing different interests. Asking whether “humanity” should pause artificial intelligence is therefore rather like asking whether humanity should abolish war. The answer might be yes without providing the slightest clue about how it could actually be achieved.

Indeed the danger of concentrating too heavily upon the concept of a pause is that it may distract attention from institutions we genuinely can build. Governments can require frontier laboratories to conduct rigorous evaluations before deployment. They can impose cybersecurity requirements upon companies possessing extraordinarily powerful models. They can create liability for reckless deployment. They can establish reporting obligations for catastrophic incidents. They can regulate the use of AI in weapons systems, critical infrastructure and public administration. They can require meaningful human responsibility for decisions involving coercive state power. They can fund alignment, interpretability and safety research. Most importantly, Washington and Beijing can begin constructing channels through which each understands what the other is doing and through which unexpected AI incidents do not automatically escalate into geopolitical crises.

There is an additional philosophical question. Perhaps we are thinking about the problem incorrectly when we ask whether artificial intelligence should be stopped. Human beings have always invented tools that subsequently altered the societies that invented them. Writing changed memory. Printing changed religion and political authority. Industrialisation changed family structures and cities. Nuclear weapons changed international relations. The internet changed privacy, journalism, commerce and political debate. We did not master these technologies before adopting them. We adapted ourselves around them, sometimes painfully and often after making serious mistakes.

Artificial intelligence may prove to be the most dramatic example of this recurring pattern because it is a technology concerned with intelligence itself. We are constructing machines that participate in precisely the activity we have historically regarded as distinguishing human beings from other creatures: reasoning about the world and deciding what to do. That is why the instinct to pause is understandable. We would like time to think before building machines capable of thinking increasingly well for us. Unfortunately the paradox is that the more valuable artificial intelligence becomes, the less willing anyone will be to stop developing it.

The tentative conclusion must therefore be uncomfortable. An AI pause might be desirable under certain circumstances. A short and genuinely reciprocal moratorium at the technological frontier could give regulators and safety researchers valuable time. Specific restraints upon exceptionally dangerous capabilities may be both possible and essential. The United States and China should certainly discuss AI safety, establish confidence-building measures and explore verification based upon the physical infrastructure of frontier computing. There is no virtue in technological recklessness merely because technological competition exists.

Yet a comprehensive pause is probably ultimately impossible. Knowledge cannot easily be contained once discovered. Computing becomes cheaper. Algorithms spread. Countries compete. Companies compete within countries. Military establishments fear strategic surprise. Researchers move across borders. Open models distribute capabilities beyond the organisations that created them. Above all, nobody can confidently know that everyone else has stopped. The same technological power that makes artificial intelligence important enough to pause also makes possessing superior artificial intelligence too valuable for states to risk falling behind.

The real choice facing humanity may therefore not be between developing artificial intelligence and stopping it. It may be between developing it with institutions capable of imposing some degree of order upon the process and developing it chaotically. The first possibility requires regulation, international diplomacy, technical safety research and an unusual degree of cooperation between geopolitical adversaries. The second requires nothing at all, because competition will produce it automatically.

Perhaps the most useful purpose served by proposals for an AI pause is therefore not that they will ever actually stop artificial intelligence. It is that they force us to confront the extraordinary character of what we are building. The United States and China may never agree to stop the race and, even if they did, they might discover that the rest of the world eventually made their agreement irrelevant. But they might agree upon some rules of the road. They might identify particular boundaries that neither wishes to cross. They might create institutions for managing accidents, misunderstandings and dangerous capabilities. They might recognise that winning a technological race is of limited value if the consequences of victory cannot be controlled.

Artificial intelligence is probably not going to pause for humanity. The more realistic task is for humanity to learn how to live with a technology that will continue moving whether or not our political institutions are ready. That is a less comforting conclusion than the idea of collectively pressing a stop button. It is also probably closer to the world in which we actually live.