Last month, Chinese artificial intelligence lab Moonshot AI introduced a new model that delivered near cutting-edge performance at a fraction of the cost of leading U.S. systems. The lab also announced plans to release an open-weight version, allowing users worldwide to download and operate the model. This development, coupled with a letter from a coalition of AI and technology firms defending open-weight models as crucial to American innovation, has reignited debates over the merits of open versus closed AI frameworks.

However, experts and policymakers argue that the primary concern is not whether models are open or closed but whether their creation stems from genuine innovation or unauthorized appropriation. Michael Kratsios, director of the White House Office of Science and Technology Policy, emphasized in a recent memo that allegations against Chinese AI firms involve industrial-scale intellectual property theft rather than open-weight AI itself.

The distinction centers on a practice known as distillation, where AI labs use outputs from powerful models to train smaller, more efficient versions. While this is standard when applied to one’s own models, employing deceptive tactics—such as fake accounts and anonymized connections—to extract proprietary knowledge from competitors’ finished models is considered theft. Companies like Anthropic and OpenAI have brought forward accusations against Chinese labs including DeepSeek, Moonshot AI, and MiniMax, alleging the use of thousands of fraudulent accounts to illicitly access their models, with estimates reaching millions of unauthorized interactions.

Anthropic’s CEO, Dario Amodei, has clarified that the company does not oppose open-weight AI in principle but is concerned about unauthorized large-scale distillation supported by the Chinese government. The White House has similarly condemned Moonshot AI for allegedly distilling Anthropic’s Fable model to develop its own Kimi K3 system.

While early Western AI models utilized extensive amounts of public and scraped data—a practice still under legal scrutiny—there is a clear difference between that approach and deliberately circumventing safeguards to extract a finished model’s capabilities. The pattern highlighted by U.S. officials aligns with other instances where Chinese firms have been accused of technology theft followed by aggressive market undercutting, such as the cases involving Huawei and Sinovel.

Defenders of China’s AI strategy argue that freely distributing a model does not constitute trade dumping since no price discrimination is involved. Yet critics contend that if a model is built by misappropriating another company’s research, low costs result from the initial investment borne by others, akin to counterfeit goods.

The implications extend beyond individual companies to broader sectors including government, healthcare, education, and national defense. As these Chinese models become embedded in widely used applications, concerns grow over increased dependency reminiscent of the controversies surrounding Huawei’s 5G infrastructure.

Chinese President Xi Jinping has described open-source AI as an opportunity to influence global technology development and reduce reliance on U.S. innovation. Beijing views AI leadership as a means to shape foundational technological frameworks critical to future economic and societal structures.

Ultimately, experts assert that open-weight AI models developed independently merit equal legitimacy with proprietary innovations shared by creators. They advocate for policies focusing on transparency regarding model provenance, trusted certification standards for systems in critical roles, and accountability for using models with unclear origins. Both open and closed AI models should compete on an equal legal and ethical footing, ensuring progress arises from genuine innovation rather than intellectual property theft.