Advances in artificial intelligence could upend the global balance of power. A state—or coalition—that first fields a decisive-capability AGI (an AI able to dominate militarily, economically, or in cyberspace) would pose an existential threat to all others, whether through coercion or accidental loss of control. Because that tipping point is hard to detect, rivals have incentives to act pre-emptively. To defuse this instability, we should steer development toward a shared AI programme operating under verifiable, multilateral oversight.
A minimum viable argument
The moment AI becomes a decisive capability will likely be recognised only in hindsight[1].
Faced with potential permanent subordination, states may contemplate military—including nuclear—options before an adversary crosses that threshold.
Uncertain action thresholds heighten pressure for pre-emptive strikes and erode confidence in traditional deterrence.
The credible alternative is to redirect competition toward verifiable international cooperation, for which many concrete tools already exist[2].
Centralised, transparent development under oversight maximises benefits and contains risks; stability improves as safety assurances and shared gains are iteratively strengthened.
Further considerations
Any front-runner in an AI race quickly becomes a perceived threat to those without adequate assurances; early cooperation is the cheapest antidote.
Economic gains from AI are similar whether pursued collaboratively or unilaterally, but unilateral paths carry far greater risks of conflict and loss of control.
During disputes, transparent mechanisms to slow or pause AI development will be essential to neutralise nuclear-level escalation risks.
Recommendations
Nations should set progressively stronger, achievable milestones for cooperation in AI development. Growing risks of mutually assured destruction—and the limited upside of an arms race—must be acknowledged and addressed early.
AI safety advocacy should emphasise both the disadvantages of competitive races and the advantages of international cooperation, promoting a transparent, verifiably collaborative pathway in which no rational actor finds opaque unilateral development attractive. Such a path offers a high-surplus, low-risk equilibrium.
Towards mutually assured cooperation
Advances in artificial intelligence could upend the global balance of power. A state—or coalition—that first fields a decisive-capability AGI (an AI able to dominate militarily, economically, or in cyberspace) would pose an existential threat to all others, whether through coercion or accidental loss of control. Because that tipping point is hard to detect, rivals have incentives to act pre-emptively. To defuse this instability, we should steer development toward a shared AI programme operating under verifiable, multilateral oversight.
A minimum viable argument
The moment AI becomes a decisive capability will likely be recognised only in hindsight[1].
Faced with potential permanent subordination, states may contemplate military—including nuclear—options before an adversary crosses that threshold.
Uncertain action thresholds heighten pressure for pre-emptive strikes and erode confidence in traditional deterrence.
The credible alternative is to redirect competition toward verifiable international cooperation, for which many concrete tools already exist[2].
Centralised, transparent development under oversight maximises benefits and contains risks; stability improves as safety assurances and shared gains are iteratively strengthened.
Further considerations
Any front-runner in an AI race quickly becomes a perceived threat to those without adequate assurances; early cooperation is the cheapest antidote.
Economic gains from AI are similar whether pursued collaboratively or unilaterally, but unilateral paths carry far greater risks of conflict and loss of control.
During disputes, transparent mechanisms to slow or pause AI development will be essential to neutralise nuclear-level escalation risks.
Recommendations
Nations should set progressively stronger, achievable milestones for cooperation in AI development. Growing risks of mutually assured destruction—and the limited upside of an arms race—must be acknowledged and addressed early.
AI safety advocacy should emphasise both the disadvantages of competitive races and the advantages of international cooperation, promoting a transparent, verifiably collaborative pathway in which no rational actor finds opaque unilateral development attractive. Such a path offers a high-surplus, low-risk equilibrium.
There’s No Fire Alarm for Artificial General Intelligence
For example, Effective Mitigations for Systemic Risks from General-Purpose AI