The international debate over the pace of AI development and necessary safeguards brings together two related but distinct positions: US technology leaders pressing for slower rollouts and independent checks, and Chinese state actors who reject alarmist rhetoric yet highlight different security risks. Both sides frame “AI security” in ways that reflect their priorities, complicating cooperation on shared rules.
What US tech leaders are proposing
Dario Amodei, chief executive of Anthropic, has argued for a slower pace of capability development and for stronger independent oversight, saying current safety measures do not keep up with increasingly powerful models. Amodei’s proposals also include curbs on Chinese access — for example tighter chip export controls and restrictions on using outputs of US models to train competing systems. Those steps are meant both to reduce safety risks and to preserve US technological advantage.
How China has responded
According to Bloomberg, the Chinese Ministry of Foreign Affairs rejected the alarmist tone of warnings from some US AI leaders. Spokesperson Guo Jiakun said that fearmongering, confrontation and ruthless competition hinder global governance of artificial intelligence. The ministry did not directly accept a proposal to slow the pace of development. At the same time, reporting in the Financial Times notes that China’s Ministry of State Security has issued its own warnings: Minister Chen Yixin highlighted risks from deepfakes, mass-produced misleading information, automated tools that stoke social division, as well as cyber security and protection of sensitive data.
Why the positions appear to conflict
On the surface the positions look contradictory: one side urges slowing development, the other rejects alarmist rhetoric yet still flags hazards. The reason is that “AI security” covers different objectives. For some actors it means preventing systems from providing dangerous assistance or performing unauthorized actions; for states it can also mean protecting the information environment, strategic infrastructure and political stability. Those goals can overlap but are not identical, and they call for different measures.
Two decisions that should be separated
Analytically it helps to separate two distinct policy choices: (1) what checks are required before a model is deployed; and (2) which countries can access the resources needed to develop models (for instance specialized chips or large datasets). If these are bundled into a single political package, the debate over export controls can undermine agreement on safety measures. For example, it would be possible to seek professional consensus about the need for independent audits even if parties disagree on chip trade — but that requires separating the content of oversight, the authority of auditors and how findings are used.
Practical considerations for corporate decision‑makers
For companies, assessing AI safety requires more granular information than national statements: details on model capabilities, data-handling terms and operational controls. The risk profile depends heavily on use case: a question‑answer assistant drawing on internal knowledge poses different hazards than an agent that executes actions in a business system. Access levels should therefore match the task and the potential consequences of errors, with active monitoring in operation. Procurement decisions are easier when providers disclose the tests performed, known limitations and incident response procedures — this is often more useful than broad claims about responsible development.
Where international cooperation could start
Concrete starting points for cooperation include common testing principles, clear incident definitions and verifiable commitments. These elements make it easier to craft interpretable agreements than trying to settle a single standard for the “right” speed of progress.
How AI security becomes credible
Ultimately, credibility depends on the consequences that follow unfavorable audit results: will deployments be delayed, access reduced, defects fixed, or will findings be used to justify trade restrictions? AI safety becomes accountable when there are pre‑agreed consequences if a system fails to meet expectations.
