On October 2, 2026, the Chinese version of the website of Radio France Internationale (RFI), a French international broadcasting station, published an article by a special correspondent based in Washington arguing that while the United States and China are competing for supremacy in the field of artificial intelligence (AI), they also face the common challenge of controlling increasingly sophisticated AI.

U.S. President Trump and Chinese President Xi Jinping met in Washington on September 24 and agreed to establish a dialogue framework to exchange views on the risks and benefits of AI, and a communication channel to respond to AI-related contingencies. Xi reportedly stated, "China and the U.S. can continue dialogue on AI, exchange views on its risks and benefits, and jointly prevent the misuse and abuse of AI," and also argued that the two countries "should learn from each other's strengths and compensate for each other's weaknesses, and should not be wary of each other."

However, the article pointed out that the core conflicts over science, technology, and security were not resolved at the meeting, and the competition over advanced semiconductors, AI models, and computing power has not subsided. Trump reportedly said in September, "Whoever wins AI is the winner," and the article describes this statement as indicative of the United States' strategic position in the AI competition. While the United States has advanced AI research institutions, cutting-edge semiconductors, and abundant funding, China is rapidly closing the technology gap and expanding global use of its low-cost, highly open AI models.

The article also highlighted that AI agents are beginning to exhibit behaviors such as deception, bypassing limitations, and hiding failures during testing. A Reuters investigation of over 200 research documents, papers, and technical reports found at least 20 studies and evaluations recorded since 2025 involving AI agents exhibiting deception, bypassing limitations, self-replication, and challenging human-defined boundaries. In one test, AI agents operating with models from China's Alibaba, DeepSeek, and Moonshot AI (月之暗面) allegedly misrepresented their capabilities to increase their chances of winning a simulated bid and continued to deceive thereafter. In another test, agents that failed to complete a task reportedly concealed their failure by falsifying results and creating fake files. The article argues that the issue of AI safety is no longer limited to incorrect responses from models but is a question of whether humans can effectively maintain control over AI systems as they become more autonomous.

The article introduced an essay contributed by Gina Raimondo, former Commerce Secretary in the Biden administration, to Time magazine. Raimondo argued, "I think AI companies are right to want to slow down the pace of developing cutting-edge technology. But even while slowing down, we must continue to lead China." She stated, "While we need to beat China in the technology itself, we also need to cooperate with China on the risks posed by AI. Competition and cooperation often conflict, but this time the stakes are too high to choose only one."

Regarding semiconductors, the United States has restricted exports of cutting-edge AI semiconductors and related equipment to China since 2022. According to the article, the U.S. has recently been focusing on whether Chinese companies are accessing advanced computing capabilities through overseas data centers, and the U.S. Congress is considering restricting methods for Chinese companies to use U.S.-made semiconductors via overseas cloud services. Raimondo argued that export controls should continue to be updated to prevent remote use through data centers and circumvention of export restrictions on controlled technologies, and called for further restrictions on the supply of semiconductor manufacturing equipment to China in coordination with major equipment producing countries such as the Netherlands, Japan, and Germany.

Another point of contention is "model distillation." This method, common in the AI industry, involves training a smaller model on the output of a larger model to achieve some capabilities at lower cost. However, the U.S. government and some AI companies accuse multiple Chinese AI companies of acquiring large-scale outputs from U.S. models and using them to train their own models. Raimondo cited data from Anthropic, stating that the company classified over 16 million interactions with Claude as activities related to model distillation, with over 3 million of these believed to be related to Chinese AI research institutions such as Moonshot AI (月之暗面). In response, China denies these accusations, stating that the U.S. claims are baseless.

Another characteristic of China's AI industry is the development of "open-weight" models, which developers can download, modify, and operate themselves. According to OpenRouter data reported by CNN, the global usage share of Chinese AI models has increased from less than 15% to over 54% in the past year. It is believed that some companies are choosing low-cost, self-operable Chinese models over expensive U.S. models, even if the latter have superior capabilities. The article discusses that the definition of "who wins the AI competition" could vary depending on whether performance, market share, number of users, or corporate profits are used as criteria. However, it also introduced an analysis suggesting that Beijing may restrict the public release of AI with capabilities that could lead to cyberattacks, from a security perspective.

Amidst a lack of mutual trust, cooperation in AI safety is expected to remain limited. Raimondo argued that while the U.S. should not blindly trust China, dialogue is not impossible, and a mechanism to verify compliance with agreements should be established. Citing the example of U.S.-Soviet nuclear arms control during the Cold War, she pointed out that even rival nations can reduce risks through long-term dialogue and verification systems. Areas where cooperation could be relatively easier include preventing the misuse of AI by cyber criminals and terrorist organizations, and stopping high-performance models from attacking critical infrastructure or aiding in biological weapon development. Some security experts are also advocating for the establishment of a "U.S.-China AI hotline" to prevent AI-induced cyber incidents from being misinterpreted as intentional attacks. (Translation/Editing by Koun)