OpenAI, Anthropic, Google and other AI companies may agree on safety standards or create a framework to manage the risks of advanced AI.
But they are unlikely to slow the underlying race for long.
A smaller laboratory could release a better model. An open-source group could produce a cheaper model. A company in another country could move faster. Investors would fund whoever appeared capable of taking the lead.
This is where the prisoner’s dilemma applies to AI.
Every company may prefer a world where all competitors develop AI carefully and safely. But each company also has an incentive to continue improving its own models while hoping that competitors slow down.
If everyone slows down, the industry benefits collectively. If one company slows down while others continue, that company may lose customers, developers, investment and market share.
China makes a coordinated slowdown even harder. The problem is no longer limited to competition between private companies. Every major country now wants its own sovereign AI models and computing capabilities.
China has made rapid progress in large language models and open AI systems. Researchers have argued that restrictions intended to preserve American leadership may have encouraged China to build more self-reliant and open ecosystems instead. Research paper
No country will want to slow down if it believes another country may use that opportunity to move ahead. Once AI becomes linked to economic growth, military capability and national influence, stopping the race becomes even less realistic.
There can be regulation and common safety standards. Companies can be required to test models, report incidents and improve security controls.
But safety may not keep pace with capability. The definition of safe and acceptable behaviour will also change over time. A model behaviour that appears harmless today may become a serious problem when future systems gain more autonomy, access to tools or influence over important decisions.
Safety concerns may therefore lead to more testing and oversight. They may not lead to slower model development.
Every country is now in a race to build its own AI capability. Once that race has begun, it cannot simply be paused by agreement.