Competition can improve safety or undermine it
The phrase “AI race” compresses several contests: companies compete for users, talent, capital, chips, and model leadership; states compete for economic growth, military advantage, and influence over standards. Competition can produce better security, redundancy, and scrutiny. It can also reward being first while making caution look like unilateral surrender.
The 2026 International AI Safety Report describes a speed-versus-safety tradeoff: when safeguards are costly or delay deployment, actors expecting an early-mover advantage may invest less in them. The report treats this as an incentive and governance problem, not proof that every laboratory is secretly abandoning safety (International AI Safety Report 2026).
What is observable today
Governments openly treat AI as strategic. The United States restricts exports of certain advanced computing items and supports domestic semiconductor production. China pursues domestic capacity and controls some upstream materials and technologies. The EU regulates access to its market while investing in compute. Many other countries seek sovereign infrastructure, local-language models, and a role in standards.
Official strategy documents are evidence of priorities, not neutral forecasts. The U.S. 2025 AI Action Plan, for example, frames policy around global leadership and rapid infrastructure development (White House, July 2025). Export-control announcements state intended restrictions; they do not establish how effectively controls work after smuggling, stockpiling, substitution, and efficiency improvements.
At company level, frequent model releases and benchmark comparisons create visible pace pressure. Voluntary safety frameworks and pre-deployment evaluations also show that competition can occur on trust and security. To assess actual corner-cutting, investigators need evidence such as shortened testing, ignored evaluation thresholds, concealed incidents, weaker access controls, or retaliation against internal dissent—not merely a fast release cadence.
The classic coordination problem
Suppose two actors both prefer a world in which everyone spends more time testing. If each believes the other will deploy immediately, delaying alone may impose commercial or strategic costs without changing the overall risk. Mutual restraint can leave both better off, but only if commitments are credible and violations detectable.
AI differs from historical arms control. Software and knowledge diffuse more easily than missiles; civilian and military uses overlap; relevant capabilities are difficult to define; private companies are central actors; and evaluation results can be noisy. Compute infrastructure is more observable than algorithms, yet algorithmic efficiency can change how much capability a given amount of compute produces.
The analogy is still useful for verification design: declarations, inspections, technical monitoring, incident hotlines, challenge procedures, and reciprocal consequences. It becomes misleading when it implies that model counts or compute map cleanly onto warheads.
National security can widen the risk frame
Advanced AI may support intelligence analysis, cyber operations, logistics, weapons development, autonomous systems, and influence campaigns. It may also improve defensive cyber work, verification, scientific cooperation, and crisis communication. Net effects are uncertain and domain-specific.
Integrating immature systems into nuclear warning, command, or other time-critical decisions creates special concern because confident errors and automation bias could compress decision time. That does not mean current chatbots control nuclear arsenals. It supports clear human authority, restricted permissions, rigorous validation, and bilateral communication about dangerous uses.
Secrecy creates a tension. States need to protect operational information, but excessive classification can prevent outside evaluation and public accountability. Independent oversight with secure access can provide scrutiny without publishing exploitable details.
Race language can become self-fulfilling
Leaders who repeatedly say that any delay guarantees defeat may narrow the political space for reciprocal restraint. Claims that a rival is inevitably reckless can justify one’s own acceleration. Threat inflation may also channel subsidies toward incumbents and national champions.
The opposite error is assuming cooperation language means interests are aligned. Declarations can coexist with competition, espionage, and military planning. Analysts should distinguish rhetoric, formal obligation, budget, implementation, and verified behavior.
Distribution matters globally
An exclusive focus on the United States and China leaves much of humanity as rule-takers. Countries in Africa, Latin America, South and Southeast Asia, the Middle East, and the Pacific may bear labor, information, extraction, and security effects without controlling leading models. Broad participation is not symbolic: language coverage, energy burdens, development priorities, and acceptable risk differ across regions.
Compute concentration can create dependency on foreign cloud providers. Sovereign AI initiatives may improve resilience and cultural representation, yet can also enable domestic surveillance or inefficient subsidy races. Governance should evaluate who controls systems and under what law, not assume “sovereign” means publicly accountable.
Practical off-ramps
Useful measures can be pursued before a comprehensive treaty:
- common definitions and evaluation protocols for dangerous capabilities;
- protected incident reporting and rapid cross-border notification;
- shared scientific assessments with documented disagreement;
- compute and supply-chain monitoring subject to privacy and trade safeguards;
- reciprocal evaluation access at defined capability thresholds;
- commitments to keep human control over specified high-consequence decisions;
- crisis hotlines and norms against targeting AI-safety infrastructure;
- support for countries otherwise excluded from testing and standards.
These mechanisms should have trigger conditions and audit procedures. A voluntary pledge is evidence of intent; implementation reports and independent tests are stronger evidence of conduct.
How AGI changes the scenario
No agreed evaluation has established AGI, so an “AGI race” remains a projection. If systems could automate AI research, cyber operations, strategic planning, and persuasion, actors might expect a temporary lead to compound rapidly. That belief—whether accurate or not—could increase pressure to deploy systems that are less understood or more autonomous.
Several uncertainties cut the other way. Gains may diffuse quickly; bottlenecks may remain in chips, energy, institutions, and physical operations; safety research may scale with capability; or leading actors may recognize shared vulnerability and coordinate. Forecasts must state these alternatives rather than treat runaway first-mover advantage as a law.
What readers should watch
Track operational indicators, not only summit language: evaluation disclosure, serious incidents, chip flows, data-center construction, military doctrine, procurement safeguards, whistleblower cases, and whether laws are enforced. Ask whether an initiative reduces shared risk or mainly shifts advantage. Check dates because export rules and corporate frameworks change quickly.
The responsible conclusion is conditional. Strategic competition creates a credible incentive to trade safety margin for speed, and current policy explicitly frames AI leadership as a national goal. It has not made catastrophe inevitable, nor shown that cooperation is impossible. The best response is reciprocal, verifiable restraint in the highest-risk areas while preserving beneficial competition in safety, access, and applications.