A senior Chinese state researcher backing shared safety tests gives the planned US-China AI safety dialogue a concrete proposal, as Beijing weighs export curbs on cyber-capable models.
Chen Yunji, a co-founder of Chinese AI chip designer Cambricon, has called for a universal "ethics checker" that every leading AI model, Chinese or American, would have to pass. Speaking on an AI safety panel on October 30, the opening day of the World Laureates Forum in Shanghai’s Lingang area, he named systems from DeepSeek, ByteDance and OpenAI as the kind of models it should cover, the South China Morning Post reported. "All human beings and nations would need to agree on this ethics checker … to prevent these large language models from doing harm to human beings in the future," he said.
The weight of the call comes from who made it. Chen is deputy director of the Chinese Academy of Sciences’ Institute of Computing Technology and heads its State Key Laboratory of Processors (处理器芯片全国重点实验室), according to Jiemian, whose report was carried by Sina Tech. He founded Cambricon with his brother Chen Tianshi in 2016, and the company is now worth more than RMB600 billion, SCMP reported.
What the proposal lacks is machinery. Chen did not say who would design the test, who would run it, or how countries would agree on it, and the Chinese-language reports of the forum do not mention it at all.
Chen expects AI that improves itself, and wants AI to police it
Chen paired the proposal with a forecast. He warned that AI systems may soon be able to improve themselves autonomously, a process known as recursive self-improvement, and said the opacity of AI decision-making could make it impossible to prevent all harm, SCMP reported. He said he had "faith" that the process could be made safe by using AI at each stage to monitor the more powerful systems that follow.
He placed chips at the center of that loop. "We use chips to accelerate AI, and at the same time we use AI to design chips," Chen said, according to Shanghai Observer, in a report republished by Tencent News. "Further on, these two things will form a cycle: AI designs better chips, and better chips train smarter AI." Jiemian reported that he named this the most important breakthrough direction of the next ten years.
The timing ties his warning to live security concerns. On October 29, cybersecurity firm CrowdStrike said an unidentified group had used an open-source DeepSeek model and another model to attack more than 200 organizations worldwide since October 17, SCMP reported. In October, Beijing proposed export controls on advanced AI systems, with extra restrictions for models that cross cyber-hacking thresholds.
His lab already lets AI draft and check processor designs
Chen’s argument rests on work his own laboratory is doing. Because a finished chip must be close to 100% accurate, his team does not expect AI to get a design right in one pass. The AI drafts a design, checks it and corrects it automatically, hundreds of times a second, Shanghai Observer reported.
The first result, Qimeng 1 in 2021, performed roughly like Intel’s 486 processor from the 1990s. Chen’s lab has had AI design a processor, Qimeng 2, that reached the level of mainstream phone CPUs from more than a decade earlier. It appeared in 2025, a jump of several generations in four years, though still well behind current commercial chips.
The research line is long. Chen’s team began studying AI-specific chips in 2008 and created a dedicated instruction set for AI workloads, and Huawei’s first-generation AI chip later used the team’s design, Shanghai Observer reported. That lineage also produced Cambricon. At a roundtable, Chen added that statistics-based large models struggle to reach full accuracy and will need physical and logical rules built in, Jiemian reported.
A shared test needs both governments, and they meet twice more this year
Chen was not the only speaker urging cooperation. Martin Hellman, the American cryptologist and Turing Award winner, said there was "no way to stop AI in the current world," given the economic and military incentives, and argued that China and the US needed a world in which cooperation on AI safety stayed possible, SCMP reported. Others were more measured: Andre Geim, the 2010 Nobel physics laureate, called AI a calculating machine without perception whose impact depends on how people use it, China News Service reported in a story carried by Sina Finance.
The forum drew 33 winners of top science prizes, including seven Nobel laureates and three Turing Award recipients, IT Home reported before it opened. It produced no joint statement on Chen’s idea.
The channel where a shared test could take shape already exists on paper. China and the US agreed in September to set up a bilateral AI safety dialogue and an incident-notification mechanism, though its scope has not been disclosed, and Xi Jinping and Donald Trump are due to meet in Shenzhen and Miami before the year ends, SCMP reported. Any agreement there on common evaluation methods for frontier models would be the first sign that Chen’s checker has political backing.