Debate grows over whether the US and China can agree on AI safety rules

In short: As the US and China compete to build stronger AI, experts are questioning whether the two countries can realistically agree on shared safety rules.

What's going on

A Financial Times opinion column looks at the idea of “AI arms control”, meaning international rules to reduce the risks of advanced AI systems. The author argues that competition between Washington and Beijing makes a meaningful deal hard.

One reason is pace. New AI models appear often, and China’s systems are improving quickly. The column points to Z.ai’s GLM 5.2 model and notes claims that some Chinese “open-weight” models (ones where parts of the model are shared publicly, like sharing a recipe instead of only selling the finished meal) are approaching the quality of leading US models.

The piece also says it is getting harder to measure which model is “better.” Common tests can be misleading, especially if a model has seen similar questions during training. Capability also depends on speed and cost, not just raw performance.

On possible safety rules, the column argues that cyber security should be taken “off the table” because both sides already use AI for hacking and defense. It suggests a more realistic focus could be biosecurity, meaning whether AI could help people make dangerous pathogens. Even there, trust is a problem.

What to watch

The biggest obstacle may be verification. During the Cold War, countries could at least count missile sites from space. With AI, the most sensitive parts are hidden in code and internal systems, and neither side is likely to open them up. Without “trust but verify,” the author suggests both countries may keep racing instead of signing a major agreement.

Source: Financial Times