Introduction
On April 23, Michael Kratsios, Director of the White House Office of Science and Technology Policy, criticized China-based actors for allegedly stealing the intellectual property of U.S. AI research institutions on an “industrial scale.” The U.S. side claims that the capabilities and information of advanced AI models are being systematically extracted. In response, Guo Jiakun, spokesperson for China’s Ministry of Foreign Affairs, rejected these claims as “groundless,” arguing that they are a smear against the achievements of China’s AI industry. This exchange is not merely one scene in the broader U.S.–China confrontation. It is an event that forces us to reconsider what “intellectual property” means in the age of AI.
The Core Problem Is That “What Was Stolen” Is Hard to See
Traditional intellectual property infringement was relatively easy to imagine. Examples include using patented technology without permission, copying source code, or reusing design drawings. In the era of generative AI, however, the subject matter that must be protected becomes far more ambiguous.
If the code or training data of an AI model itself is stolen, the issue is relatively straightforward. However, what is being treated as problematic in this case is conduct such as sending a large number of questions to an AI model and using its outputs or response tendencies to train another model. Reports describe this method as “distillation,” explaining it as a technique for transferring the capabilities of a powerful AI model to a smaller model.
The difficult question here is how far the “capabilities” or “behavior” of AI should be protected as intellectual property. The boundary between a human learning by using a public service and an organization systematically making large-scale access to build a competing model cannot be easily drawn, either technically or legally.
U.S. Concerns Are Not Limited to Technology Leakage
The concerns on the U.S. side are not simply about “the profits of domestic companies being taken away.” AI has become a foundational technology directly connected to national security, including in the fields of military affairs, cybersecurity, semiconductors, healthcare, and finance. For this reason, the transfer of advanced AI model capabilities to a competing country is viewed not only as an issue of industrial competitiveness, but also as a national security concern.
In particular, the development of foundation models requires enormous amounts of capital, computing resources, talent, data, and accumulated research. If companies in other countries can imitate those achievements in a short period of time, this creates a major disadvantage for the companies and countries that made the original development investments. Behind the U.S. use of the strong phrase “industrial scale” is a sense of alarm that its first-mover advantage in the AI development race could be rapidly lost.
China Also Has Reasons to Push Back
On the other hand, it is also natural that China would push back. In the AI field, Chinese companies and research institutions are also advancing their own research and development. If everything is regarded as “theft,” China may perceive this as a denial of its own technological progress.
Moreover, in AI development, various forms of information circulate across borders, including published papers, open-source models, public benchmarks, API use, and synthetic data. Even from the outside, it is difficult to determine where legitimate research and development ends and improper acquisition of intellectual property begins. The Associated Press has also reported, citing experts, that distinguishing unauthorized use from legitimate use is complex.
In other words, this conflict cannot be fully understood through the simple framework of “the U.S. is right” or “China is right.” The essential problem is that the rules surrounding AI intellectual property have not kept pace with technological evolution.
In the AI Era, Protecting IP by “Keeping It Secret” Is Not Enough
Until now, corporate intellectual property strategies have mainly centered on filing patents, managing information as trade secrets, protecting works through copyright, and imposing contractual restrictions. In the case of AI models, however, once they are provided externally as services, users can access their outputs.
Of course, it is possible to prohibit the training of competing models through terms of use. However, actually detecting violations, proving them, and pursuing responsibility across borders is not easy. This is because AI model outputs change each time, and the pathways by which outputs are used for training are difficult to see.
Going forward, AI companies will need to focus on monitoring model usage, detecting suspicious large-scale access, tracking characteristics contained in outputs, tightening API usage conditions, and strengthening international enforcement cooperation. Intellectual property protection is likely to become integrated not only with patent applications and contracts, but also with cybersecurity and data governance.
Lessons for Japanese Companies
This news is not only relevant to major AI companies in the United States and China. It also offers important implications for Japanese companies.
First, when using AI, companies need to organize how they are using the outputs of other companies’ AI systems. If AI outputs are accumulated in large quantities and used for training internal models or developing services, this may conflict with terms of use or contractual restrictions.
Second, when protecting a company’s own AI-related technology, patents alone are not sufficient. It becomes important to manage assets that are difficult to see from the outside, such as training data, prompt design, evaluation data, tuning methods, operation logs, and know-how for improving the response quality of models.
Third, in an AI-era intellectual property strategy, companies need to design in advance what to disclose, what to keep confidential, and what to restrict by contract. Responding after technology has leaked is too late. It is necessary to build mechanisms for intellectual property protection into the service provision stage itself.
Conclusion
This U.S.–China confrontation is both a struggle for dominance over AI and an indication that the very concept of intellectual property is changing. The value of AI resides not only in source code and training data, but also in the capabilities acquired by models, the tendencies of their responses, and the quality of their reasoning.
However, how far these elements can be legally protected, how improper acquisition can be proven, and where the boundary with legitimate research and development should be drawn have not yet been sufficiently clarified.
AI-era intellectual property protection is moving beyond the stage of simply “preventing theft” and into a stage of “monitoring how AI is used and protecting it through a combination of contracts, technology, and institutions.” This news can be seen not only as a sign of intensifying competition in AI development, but also as an event that demands a redesign of intellectual property strategy itself.