The coming disorder over artificial general intelligence and how to mitigate its impact

Public and private decision-makers often describe artificial general intelligence, or AGI, as a race. The frame is comforting – and wrong.

The AI ‘race’ is not about crossing a finish line of humanlike capabilities, much less god-like intelligence. It is about the spread of machine intelligence through companies, governments and public services before political institutions are ready for it.

Artificial intelligence is already diffusing through the global economy faster than institutions can adapt, while its gains are poised to concentrate among those who own the compute and distribution channels through which intelligence will be deployed. The International Monetary Fund (IMF) has warned that generative AI could both raise productivity and widen inequality.

In this light, the rhetoric of an AGI race serves the interests of frontier labs and hyperscalers seeking to justify vast investment in chips, data centres and energy infrastructure. It is less useful for governments and societies left to manage labour displacement, increasingly sophisticated cyber threats and degraded sovereignty.

The question is not who reaches AGI first, but who controls the physical and digital systems through which machine intelligence becomes ordinary?

The familiar story is that the US and China are locked in a contest for technological supremacy. Washington has frontier labs and chip dominance, while Beijing has scale, industrial depth and an incentive to make deployment cheaper. The two powers are often described as separated by months, each racing to deny the other a decisive advantage. But this story doesn’t adequately reflect the problem governments face.

Artificial general intelligence remains poorly defined, and predictions about its timing often obscure more than they clarify. The US has a lead in private AI investment, but Chinese models have narrowed performance gaps across key benchmarks related to cost, speed and scalability.

The claim that one country or company can “win” falls apart once one asks what victory would mean. The most capable model, the largest enterprise market, the highest productivity gains or the most advanced surveillance system each confer a different kind of power.

The rhetoric of winning nevertheless sustains the case for extraordinary investment by hyperscalers and their shareholders, which needs a story large enough to justify it. That framing turns commercial acceleration into strategic necessity and shifts attention away from the societies carrying the greatest risks.

The disruption will be felt through the spread of general machine intelligence into the routines of the state, markets and everyday life. It is already moving into welfare systems, corporate operations and military planning. If a formal AGI threshold is ever declared, the political economy of intelligence will have already changed.

The race metaphor also hides the emerging geography of AI power. At the top are the US and China. The US follows a concentrated, capital-intensive path, while China pushes towards efficiency and domestic substitution.

US chip restrictions can slow Beijing, but they also force it to innovate around scarcity. A system optimized for lower cost and wider deployment can become geopolitically significant. Analysis of DeepSeek’s low-cost reasoning model concluded that it remained competitive despite being developed under US graphics processing unit export constraints.

Below them are middle power states trying to build enough AI capacity to avoid dependence on foreign models, chips and clouds. Canada, France, India, Japan, Saudi Arabia, South Korea, Singapore, the UAE and the UK have the capital, talent or strategic motive to make a serious attempt.

Other countries will seek narrower autonomy in public services, defence and regulated industries. Most will not build frontier models and many will still need more compute and deeper capital markets to retain control over critical functions.

The third tier includes most other states that will use what is available. Open-source models and specialized applications will enable governments and companies to adapt increasingly general systems to local needs. They may create some autonomy, but will also deepen dependence on architectures designed elsewhere.

Paradoxically, the biggest near-term gains may accrue to those willing to exploit open source and edge models rather than frontier systems. Many valuable applications do not require a trillion-parameter model in a distant data centre. Smaller systems can transform health diagnostics and fraud detection while running closer to users, inside companies and within public agencies.

This matters because general capability will arrive through agents and domain-specific systems that act across real environments, as well as large frontier models. Diffusion increases capability and exposure at the same time.

Open models and edge systems widen the attack surface and reduce visibility into downstream uses. The same systems that improve public administration can extend surveillance or automate reconnaissance.

The current focus also misses the compounding nature of the risks, as governments still treat AI and AGI risks in isolation.

National security will be altered by autonomous weapons and synthetic propaganda. Markets will be reordered as capital substitutes for labour. Sovereignty itself will be tested as countries discover that formal independence means less when the intelligence layer is owned abroad.

The dominant AGI investment thesis assumes returns will flow to the owners of compute and distribution. That may comfort shareholders, but it is less reassuring for workers and taxpayers.

If machine intelligence increases productivity while weakening labour income and concentrating rents, political and social destabilization could arrive quickly.

Few societies are preparing at the scale the disruption requires. Governments are still speaking the language of innovation strategies and voluntary safeguards. Companies are racing because investors reward speed. Citizens are being asked to trust institutions that have already lost much of their authority.

The labour shock is the most immediate political danger. AI does not have to eliminate work wholesale to destabilize societies. A compression of wages and career ladders across clerical, customer-service and software-adjacent roles could damage household income and weaken the link between education, effort and reward on which many democratic societies rely.

Democracies face a particular problem because their institutions struggle to plan over long horizons. Regulatory systems are fragmented, and publics are already suspicious of experts, let alone corporate and political leaders.

China’s centralized model may be better suited to directing capital and mandating adoption. But that is not the same as social resilience or innovation quality. It is a warning about democratic capacity, not an endorsement of centralized control.

The default trajectory is disorder, not abundance. The US will keep pushing the frontier to avoid losing technological advantage. China will diffuse capability because it sees a path around US dominance.

Middle powers will seek practical autonomy in order to diminish their reliance on foreign chips, clouds and models. Companies will integrate increasingly capable agents faster than governments can supervise them.

Avoiding that default requires a more serious politics of AGI. Leaders must treat machine intelligence as a matter of national security, state capacity and public legitimacy, not as an innovation agenda.

Sovereign-AI strategies should focus on practical autonomy over critical systems, while labour policy needs to move ahead of displacement. Security governance must extend beyond voluntary pledges and narrow model evaluations.

The hardest task is rebuilding collective capacity in low-trust societies. AGI governance will require cooperation between states and companies, and rival powers. It will also require public institutions to act before crises emerge.

Governments must reframe AI as critical infrastructure rather than commercial software, with serious attention to compute access, public data systems and evaluation capacity. Labour disruption should be treated as a fiscal and political problem, while security governance should move from voluntary pledges to enforceable obligations.

Political leaders need sustained engagement with what general machine intelligence can do. That requires serious dialogue not only with technologists and social scientists, but with the wider public, since democratic legitimacy depends on bringing citizens into decisions about how machine intelligence is deployed.

For the vast majority of countries, a credible sovereign AGI strategy does not mean building a prestige frontier model. It means securing enough computational power and trusted public data to govern advanced models in critical sectors.

Labour-market preparation needs to move much faster. Governments should identify highly exposed occupations and redesign training systems before displacement becomes explosive.

Security governance must become harder-edged. Cyber and defence applications require mandatory testing, incident reporting and enforceable standards. Voluntary commitments will not hold where incentives reward speed.

Fiscal policy will also have to change. AI taxes and robot levies should be examined as ways to fund social adaptation. Universal basic income may become necessary if displacement becomes acute enough, although redistribution alone cannot replace work as a source of dignity.

AI will generate large productivity gains but who captures them and whether public institutions can act before disruption hardens into distrust remain open questions. If governments continue to treat AI as a race to be won, they will miss the harder task: building societies resilient enough to absorb intelligence they do not fully control.