It is now well-documented that AI developers know more about the risks of their systems than they say publicly. Internal safety teams at major AI labs have documented significant concerns. Researchers who leave those labs describe internal debates about whether current systems are being deployed too quickly. The safety concerns are not unknown — they are known and, in many cases, acted on partially while development continues.
This is not primarily a story about bad actors. It is a story about a coordination problem in which even safety-conscious actors face structural incentives that push them toward faster development and deployment than their own risk assessments would recommend. Understanding those incentives — the AI race dynamics — is essential to understanding why individual organizations' commitments to safety are insufficient and what kinds of governance structures are actually required.
The structure of the race
The basic structure of AI race dynamics is straightforward. Multiple actors — companies within a country, countries in the world — are competing to develop the most capable AI systems and deploy them first. Investment in safety takes time and resources that could alternatively be used to accelerate capability development. Any actor who slows down to invest seriously in safety falls behind actors who do not. The actor who wins the race is, by structural incentive, likely to be the one who traded most aggressively against safety.
This creates a dynamic in which safety-conscious actors face a choice between two bad options. They can maintain their safety standards and fall behind — resulting in the frontier being shaped by less safety-conscious actors. Or they can reduce their safety investment to remain competitive — resulting in their own standards degrading. Either way, competitive pressure systematically produces outcomes worse than any individual actor would prefer if they could coordinate.
A safety-conscious lab that slows down unilaterally does not produce a slower overall pace of AI development. It produces a faster overall pace dominated by the labs that moved fastest. Voluntary safety investment only improves overall safety if it is coordinated across all major actors simultaneously — which is precisely what market competition prevents.
The national dimension
Lab-level race dynamics are in principle addressable through domestic regulation: a government can impose safety requirements on all AI labs within its jurisdiction simultaneously, eliminating the competitive disadvantage of safety investment within that jurisdiction. This is why domestic AI regulation matters — it can resolve the within-country coordination problem, even if it does not address the between-country problem.
National-level race dynamics are structurally harder. The US-China competition in AI is frequently framed as a strategic race with significant geopolitical consequences, and that framing produces the same incentive structure at the national level: any nation that slows down for safety falls behind nations that do not. The concern that safety regulation creates a strategic disadvantage relative to less safety-regulated competitors is a genuine coordination problem, not a bad-faith objection.
This is why purely domestic AI regulation — even stringent domestic regulation — is insufficient as a response to AI race dynamics. The international governance gap is not a technical problem about which country has better regulation. It is a structural coordination problem that requires multilateral frameworks with enforcement mechanisms, analogous to the arms control treaties that partially resolved the nuclear arms race.
Precedents for resolving race dynamics
The concern that coordination on dangerous technology races is politically impossible does not match the historical record. The US and Soviet Union, at the height of the Cold War, negotiated SALT I and SALT II — strategic arms limitation treaties that imposed binding limits on both parties' most dangerous weapons, despite deep mistrust and ongoing strategic competition. The Chemical Weapons Convention banned an entire category of weapons across most of the world's nations. The Biological Weapons Convention did the same for biological weapons. The Montreal Protocol coordinated the global phase-out of ozone-depleting substances.
None of these were achieved without difficulty, defection, and enforcement challenges. All of them represented genuine improvements over uncoordinated races. The historical precedent for international coordination on dangerous technologies is more extensive than the pessimistic framing of AI governance discussions would suggest.
Why "safety-conscious labs should win the race" is not enough
A common response to AI race dynamics concerns is that the solution is for safety-focused labs to succeed, so that the frontier is shaped by actors who care about safety. This argument has some validity: if a race is occurring regardless of safety advocates' preferences, having safety-conscious actors at the frontier is better than having safety-indifferent ones.
The argument has limits that are worth stating clearly. First, being at the frontier does not eliminate race dynamics — it makes the safety-conscious lab the one now facing competitive pressure to cut corners to stay at the frontier. Second, the competitive pressure argument can be used to justify any acceleration, because there is always a less safety-conscious competitor to invoke. Third, the argument relies on safety-conscious labs winning, but provides no mechanism to ensure this — it is a hope, not a governance structure.
Resolving AI race dynamics requires the same thing that resolved the nuclear arms race partially: coordinated limits that apply to all major actors simultaneously, with verification mechanisms to detect defection and enforcement mechanisms to deter it. That is a multilateral governance problem. The Foundation's governance proposals are structured around this requirement, not around the hope that the right labs finish first.
Common questions.
The competitive pressures that push AI developers and nations to prioritize speed over safety. The core mechanism is a coordination problem: safety investment takes time and resources that could be used for capability development, so any actor who slows down for safety falls behind actors who do not. This creates structural pressure on all actors — including safety-conscious ones — toward trading safety for speed, producing collectively worse outcomes than any individual actor would choose if they could coordinate.
Individual labs can and do make safety investments. The coordination problem is that unilateral restraint in a competitive environment does not produce a safer overall outcome — it produces the same development pace but now dominated by actors who chose not to restrain themselves. Genuine safety improvement requires coordinated restraint across all major actors simultaneously, which market competition and geopolitical rivalry systematically prevent in the absence of multilateral governance frameworks.
Yes. The framing of US-China AI competition as a strategic race creates the national-level equivalent of the lab-level coordination problem: any nation that regulates more stringently for safety creates a potential strategic disadvantage relative to nations that do not. This is why domestic AI regulation, however stringent, does not resolve race dynamics at the global level. International coordination frameworks — analogous to arms control treaties — are needed to address the between-country dimension of the problem.
Yes. The nuclear arms control treaties (SALT I, SALT II, New START) partially resolved the US-Soviet nuclear race under conditions of deep strategic rivalry. The Chemical Weapons Convention banned an entire category of weapons across most of the world's nations. The Montreal Protocol coordinated global phase-out of ozone-depleting substances. None of these was easy or perfect. All represented genuine improvements over uncoordinated races. The historical precedent for successful coordination on dangerous technology races is more robust than AI governance pessimists typically acknowledge.