Race dynamics and their effect on safety

The core tension

Safety practices — extensive red-teaming, staged deployment, pausing on concerning evaluation results — all cost time. In a competitive race where falling behind is treated as a strategic loss, every actor faces pressure to spend less of that time than they otherwise would, even when every actor would prefer a slower, safer pace collectively. This is a textbook coordination problem: safety that isn’t matched by rivals just cedes ground, so no one wants to go first.

Why this isn’t hypothetical

Export controls on advanced AI chips, national AI strategies treating frontier capability as a strategic asset, and public statements from officials in multiple countries framing AI leadership in explicitly competitive terms have all been visible since 2022–23. Yoshua Bengio’s International AI Safety Report explicitly flagged that a low-cost, high-performance competitor model accelerated the broader race in early 2025:

Proposed off-ramps

Serious governance proposals in this space tend to converge on one of two structures: verifiable, mutual restraint agreements modeled on arms-control precedent, or distributing frontier capability across multiple accountable actors within and across allied nations so that no single country’s internal caution (or lack of it) determines the pace for everyone. See International Treaties and Summits for where these efforts currently stand.

The honest state of play

As of this writing, no binding international agreement constrains the race dynamic; the Bletchley Declaration and its successors are coordination signals, not enforcement mechanisms.