On September 13, 2026, the Associated Press reported that U.S. President Trump publicly downplayed the necessity of slowing cutting-edge AI R&D, with the core reason being that the U.S. cannot give up its technological lead to China. He acknowledged that AI needs certain guardrails but did not provide specific mechanisms, summarizing the outcome of competition as: whoever wins AI wins the future. This response pushed the industry's recent proposal to "control the expansion speed of the most dangerous capabilities" directly to the level of national competition policy.
The debate isn't about whether to stop AI
Anthropic CEO Dario Amodei previously emphasized the focus of "setting a rhythm for cutting-edge capabilities" rather than freezing R&D. He suggested that independent third parties enter the model training environment for evaluation, coordinate industry actions when verifiable hazardous capability thresholds are reached, and reduce concerns about being overtaken by other participants after slowing down by individual companies or countries through international collaboration. Sam Altman of OpenAI and Elon Musk of xAI also expressed support for slowing down cutting-edge training, while Demis Hassabis of Google DeepMind believes the direction is worth discussing, but the implementation details still need improvement.
Why do these two goals clash head-on?
| Policy objectives | The main concern | Evidence is needed |
|---|---|---|
| Maintaining technological leadership | Unilateral slowing may allow competitors to catch up | Comparable metrics of computing power, talent, and model capability |
| Controlling frontier risks | Capacity expansion first, evaluation and emergency mechanisms supplemented later | Hazard capability thresholds, third-party evaluations, and accident notifications |
The real difficulty lies in the fact that neither side can rely solely on slogans. Without verifiable international coordination, companies find it hard to trust competitors to comply with restrictions simultaneously; Without public evaluations and clear triggering conditions, the "guardrail" may become a statement of principle that does not disrupt the pace of R&D.
What should businesses focus on now?
For model companies, cloud vendors, and large purchasers, the near-term impact is more likely to be in compliance and procurement terms, rather than a sudden and comprehensive shutdown. Companies should require suppliers to explain security assessments, red team scope, major capability changes, and incident response mechanisms before frontier models go live, and retain permission isolation and manual confirmation for high-risk agents. Investors should also observe whether the U.S. proposes enforceable evaluation standards and whether these standards will be linked to export controls, computing infrastructure, and government procurement.
The key to this disagreement is not the choice between "safety" and "innovation," but who measures risk, when restrictions are triggered, and whether competitors accept the same set of verifiable rules. If these questions remain unanswered, leading and safety barriers will remain mutually exclusive policy goals.
