From Silicon Valley to Washington, D.C., the technology industry is rapidly turning its attention to a single AI trend: **distillation**.

Earlier this year, Google’s Chief AI Scientist, Jeff Dean, highlighted a concept that had received little attention outside specialized AI circles: **distillation**. Speaking on a podcast, Dean explained that Google adopted AI distillation techniques while seeking ways to improve the performance of its systems without relying solely on a single, large image-recognition model. He noted that distillation enables developers to transfer the capabilities of a powerful frontier model into a smaller, more efficient model, making it significantly more capable. Just a few months later, distillation has become a major topic of discussion across Silicon Valley and Washington, D.C. Technology leaders and policymakers are debating whether the technique could pose national security risks by helping China narrow the AI gap with the United States. The debate intensified after Chinese AI company Moonshot AI launched its Kimi K3 model, which many users found to perform at a level comparable with leading commercial AI systems from Anthropic and OpenAI. Unlike many U.S. AI companies that provide access to proprietary models through paid services, Moonshot AI and several other Chinese developers release open-weight models, allowing users to download, modify, and deploy the technology independently. Some U.S. officials argue that Moonshot’s rapid progress may have been accelerated through AI distillation, alleging that the company relied on technology derived from Anthropic’s frontier Fable model. White House adviser Michael Kratsios claimed on X that Moonshot AI built an advanced internal platform capable of conducting large-scale distillation of U.S. AI models while using multiple access methods to reduce the likelihood of detection. At its core, **AI distillation** is the process of using the responses or outputs of an advanced AI model to train another model. While the technique is widely used in AI development, it has sparked debate because it can enable developers to build competing systems by leveraging the work of companies that have invested millions—or even billions—of dollars in creating cutting-edge AI models. Pukar Hamal, founder of AI security firm SecurityPal, compared the practice to a student who attends every lecture, studies the material, and completes all the assignments, only for another student to copy the finished work without putting in the same effort. As concerns over AI distillation intensified, some of the world's largest technology companies—including Nvidia, Microsoft, Meta, and Palantir—joined more than 20 other firms in signing a letter urging policymakers not to impose what they called **premature restrictions** on open-weight AI models. The companies argued that excessive regulation could reduce competition and encourage AI innovation to move overseas. In their letter, the companies described distillation as a standard AI development method that uses one model's outputs to improve, refine, or validate another model, emphasizing that it is a common and legitimate practice across the industry. The growing use of distillation has also created a policy challenge for U.S. officials, who have long expressed concerns about Chinese technology, intellectual property protection, and national security. According to Colin Shea-Blymyer, a research fellow at Georgetown University's Center for Security and Emerging Technology, policymakers are still determining how to address the issue. One possible concern is that Chinese or Russian companies could use the outputs of advanced U.S. AI systems to accelerate their own development and gain a competitive advantage. Among the signatories of the industry letter was Box CEO Aaron Levie, who argued that maintaining U.S. competitiveness requires companies to have access to the best AI technologies regardless of where they originate. He added that increasing global innovation is likely to make AI systems more affordable and efficient over time. Although much of the current debate has focused on Chinese open-weight models such as Kimi K3, AI experts note that distillation is already widely used throughout the industry. Shashi Bellamkonda, Research Director at Info-Tech Research Group, pointed to Nvidia's Llama Nemotron models as one example where distillation formed part of the training process. He described the technique as a legitimate and highly effective way to develop smaller, lower-cost AI models from the outputs of larger, more advanced systems. Anthropic has taken a different stance on AI distillation, arguing that the practice poses significant risks because it directly affects its proprietary AI models and growing commercial business. In February, the company alleged that Chinese AI firms, including DeepSeek, Moonshot, and MiniMax, carried out AI distillation on an "industrial scale" by operating approximately 24,000 fake accounts that generated around 16 million interactions with its Claude models. Anthropic, which is valued at nearly $1 trillion and is reportedly preparing for a future public listing, has described unauthorized AI distillation as a national security concern. The company argues that its AI systems include safeguards designed to prevent misuse by state and non-state actors, including attempts to develop biological weapons or conduct malicious cyberattacks. It has called for coordinated action among technology companies, policymakers, and the broader AI community to curb unauthorized use of frontier AI models. Both Anthropic and OpenAI explicitly prohibit AI distillation in their terms of service. According to Shashi Bellamkonda of Info-Tech Research Group, these restrictions reflect the companies' view that using their advanced models without permission could constitute intellectual property infringement. Despite these concerns, the rapidly rising cost of developing and operating AI systems is encouraging many organizations to seek more efficient alternatives. Pukar Hamal, founder of SecurityPal, said he would be comfortable using Chinese open-weight models such as Kimi K3 for his company's AI-powered security assessment platform, provided the software is thoroughly reviewed for potential security risks. He noted that after verifying the code and deploying it on the company's own infrastructure, the technology could offer substantial cost savings. At the same time, Anthropic and OpenAI face criticism over their own use of copyrighted material during AI training. Both companies have been sued over allegations that they used protected content without authorization. Max Pritt, an attorney at Boies Schiller Flexner representing authors in copyright cases against AI companies, argued that while the U.S. government has publicly emphasized protecting the intellectual property of technology firms, it has been far less vocal about safeguarding the rights of creators whose works were allegedly used to train AI systems without permission.