Do Open AI Models Hand America's Rivals a Weapon?

One fear now shapes much of the argument over how AI should be released: that every open-weight model an American company publishes is a gift to the country's rivals, handing China and others a free copy of hard-won capabilities. It is a serious worry, and it deserves an honest look, because the evidence points in more than one direction.

An "open-weight" model is one whose trained parameters are published, so anyone can download it, run it privately and adapt it. Meta's Llama line and a run of Chinese models have made this the fastest-moving corner of the field. The anxiety is intuitive. If the weights are free, a hostile government or a criminal group gets the same tool a well-resourced lab spent hundreds of millions to build.

What the evidence actually shows

The gap between open and closed models has narrowed sharply. A recent analysis by the nonprofit SaferAI found that GLM-5.2, an open model from China's Z.ai, trails the best American systems on sensitive cyber and biology tasks by only a few months. That is the strongest version of the concern: the frontier is no longer a moat, and the most capable open models sit close behind the leaders.

But the "gift to rivals" framing assumes the rivals need the gift. Much of the open-weight momentum now comes from China itself, where well-funded labs are shipping capable models of their own. Restricting American open releases would not remove those tools from the world. It would, the argument runs, simply cede the open ecosystem to Chinese standards. That is the case 25 companies, among them Nvidia, Microsoft, Meta, IBM and Palantir, made in a July open letter titled "Open Weights and American AI Leadership," warning that curbs would push developers toward foreign models and choke off the security research that open access makes possible.

Where Washington landed, for now

The Trump administration appears to have accepted much of that logic. Its AI security framework, finalised in early August, exempts open-weight models from a new review process and instead asks the makers of the most capable closed models to submit voluntarily to a 30-day government check before release. At the same time, officials have weighed limits on Chinese open models reaching American users, a sign that the worry has not gone away so much as changed shape.

None of this makes the fear baseless. Anthropic and other safety-focused voices continue to argue that publishing weights is irreversible, and that a genuinely dangerous capability, once released, cannot be recalled. That is true, and it is the strongest reason for caution about what gets opened and when.

The honest conclusion is that "open models are a weapon for our enemies" is too blunt to be useful. Openness carries real risks, especially at the frontier of cyber and biological capability, and those deserve scrutiny case by case. But the blanket claim ignores that the rivals in question are already building their own, and that closing off the American ecosystem might weaken it without making anyone safer. The real question is not open versus closed. It is which specific capabilities are too dangerous to hand out freely, and how to judge that before the download link goes live.