OpenAIはオープンウェイトモデルを恐れている——米国も警戒すべきか？OpenAI is scared of open-weight models. Should the US be?

Open-weight AI models, systems whose trained parameters are published so that anyone can download, run, and modify them, have become one of the most contested subjects in technology policy, and OpenAI's increasingly vocal wariness toward them is drawing scrutiny. The company has warned that the unrestricted spread of capable open-weight systems carries national-security risks, but critics argue the stance protects OpenAI's commercial position at least as much as it addresses genuine danger.

The distinction at the center of the debate matters. Open-weight models are not necessarily fully open source; developers typically release the numerical weights and permit local use and fine-tuning, but often withhold training data, code, or detailed methodology. That is different from closed models such as those OpenAI primarily sells through an API, where the weights never leave the provider's servers and access can be metered, monitored, and revoked. Prominent open-weight releases include Meta's Llama family, France-based Mistral, and Chinese efforts such as DeepSeek and Alibaba's Qwen, which have narrowed the perceived quality gap with proprietary frontier systems and intensified competitive pressure.

OpenAI's argument, echoed by some policymakers, is that once weights are public they cannot be recalled. A malicious actor can strip away safety guardrails through fine-tuning, and the model can be run offline beyond any vendor's oversight. Proponents of restriction point to potential misuse in areas such as biological or chemical weapons research, large-scale disinformation, and offensive cyber operations, and they raise concerns about advanced capabilities diffusing to strategic rivals. This framing has fed into a broader US policy conversation that already includes export controls on advanced chips and rules aimed at limiting the flow of cutting-edge AI to adversarial states.

Skeptics counter that the national-security case can double as a business strategy, a pattern sometimes described as regulatory capture, in which incumbents encourage rules that raise barriers for competitors. Open-weight models let startups, researchers, and enterprises build on strong systems without paying per-token fees or accepting a single vendor's terms, which directly threatens the economics of a closed-API business. Critics also note that many of the most capable open-weight models now come from outside the United States, meaning that restrictions on American developers may do little to stop global availability while ceding ground to overseas alternatives. Some researchers have argued that open weights actually improve safety by enabling independent scrutiny, reproducibility, and faster identification of flaws.

The situation is complicated by OpenAI's own record. The company that began as a nonprofit committed to open research has largely moved toward closed products, though it released open-weight models, marketed under the gpt-oss branding, in 2025, positioning them as a controlled contribution to the ecosystem. That history makes its cautionary messaging a target for accusations of inconsistency, and observers appear divided over how much of the current posture reflects sincere risk assessment versus competitive interest. It is likely that both motivations coexist, which is part of why the debate is difficult to resolve.

For readers tracking the policy landscape, several adjacent threads are worth watching. The definition of what counts as a dangerous capability threshold remains unsettled, and proposals have ranged from mandatory pre-release evaluations to reporting requirements for models trained above certain compute levels. The National Institute of Standards and Technology and its AI safety work, along with earlier executive actions on AI, form part of the regulatory backdrop, as do ongoing discussions in the European Union and elsewhere about how to treat general-purpose and open models. Industry groups representing open-source developers have pushed back against blanket restrictions, warning of harm to academic research and smaller firms.

The core tension is unlikely to disappear soon. Open-weight models deliver transparency, cost savings, and broad innovation, while also lowering the barrier for misuse that cannot easily be undone. Whether the United States should be as wary as OpenAI suggests depends on empirical questions about real-world risk that remain only partly answered. What is clear is that any policy will carry commercial consequences, so the arguments of interested parties, including OpenAI, warrant careful and independent evaluation rather than acceptance at face value.