Trump won’t put the brakes on AI because he wants to beat China. How big a threat is Beijing?

Last weekend, as the world worried about the possibility of unconstrained artificial intelligence wiping out humanity, the leaders of the United States and China articulated their visions for the era-defining technology.

They came from almost opposite sides of the spectrum.

Brushing aside the calls from AI scientists and executives from Anthropic , OpenAI and Google to slow the development, US President Donald Trump said what was most important was keeping the US in the lead against China in the intensifying AI superpower race.

“We’re leading China in AI…and frankly, I want to keep it that way, because whoever wins AI wins,” he said on Saturday.

On Sunday, China’s leader Xi Jinping made his own call for more international cooperation. Countries should “encourage open source, openness, collaboration and sharing,” he said in a speech in India to Global South leaders, in which he reiterated a proposal for a global AI governance pact – although he did not give details on what that framework would look like.

The juxtaposition of Trump and Xi’s messages underscores the lack of consensus and a deep trust deficit between the two superpowers in the high-stakes race to dominate AI.

The US appears to be increasingly weaponizing its dominance in the sector, seeing it as a strategic asset to be protected, and exported to friendly countries, while China, lagging in the race, is proposing a more open model.

AI will loom large at the highly anticipated Trump-Xi summit in two weeks in Washington. At their May meeting in Beijing, the two discussed AI guardrails but produced limited results, and last week the Chinese Ministry of Commerce said the two leaders had agreed to set up an intergovernmental dialogue on AI. But with their broader geopolitical rivalry, experts have low expectations for any meaningful breakthrough.

Chinese AI has trailed its American competitors for years. That’s partly thanks to US technology controls that limit Chinese companies’ access to advanced AI processors — essential for training increasingly complex AI models.

But last year Chinese AI startup DeepSeek shocked the global industry with a model that approached the performance of leading American systems using what it said were far fewer computing resources.

Since then, Chinese AI labs have iterated faster, with Alibaba, Z.AI , Minimax and most recently Moonshot unveiling similarly capable models that challenge the dominance of American frontier companies like Anthropic and OpenAI.

Alongside the advances of Chinese models, however, came ever louder US accusations of cheating. Washington, as well as OpenAI and Anthropic, have alleged that Chinese companies stole American intellectual property by training their models on the outputs of frontier US models, a process known as distillation in the industry.

Last week when US federal agencies, including the FBI, National Security Agency, and Cybersecurity and Infrastructure Security Agency, claimed Chinese AI firms are committing “industrial-scale” theft of their American counterparts.

Kylie Cooper/Reuters

In a lengthy report last week, Anthropic also detailed attempts by bad actors, including alleged Chinese government-affiliated operators, misusing its systems to develop weapons and conduct surveillance on minorities like Uyghurs in Xinjiang.

In one example, Anthropic uncovered sensitive information about Chinese surveillance attempts because queries from a user – believed to be linked with the Chinese military – sent to Moonshot’s Kimi model were rerouted to Claude, exposing Moonshot’s secret attempt to divert user requests to Claude, the report said. Other Chinese firms like DeepSeek also engaged in such a practice, it added.

Beijing has rejected the allegations, while Chinese AI labs have largely stayed silent. Moonshot declined to comment on the report to CNN last week, while DeepSeek did not respond to a request for comment.

“Spreading various threatening narratives, pursuing confrontation and malicious competition will only disrupt efforts toward global AI governance and serve the interests of none,” said Foreign Ministry spokesperson Guo Jiakun on Monday.

“All parties should work together to promote AI that is open, inclusive, accessible and beneficial to humanity,” he added.

Just as the US feels increasingly threatened by China’s rapidly advancing models, Beijing is also alarmed by the cyber capabilities of systems developed by American companies – capabilities on display this summer when agents from OpenAI and Anthropic went rogue and launched unauthorized hacks of other systems.

Beijing also sees AI safety risks differently, beyond the US understanding of existential risks, like AI agents going rogue. Besides prioritizing political control and maintaining social stability, it sees the US as a “hegemonic threat” to its own development of the technology.

In an article on Sunday, State Security Minister Chen Yixin highlighted the risks of “hegemonic practices” in a veiled jab at the US for “establishing entity lists, imposing technology controls, monopolizing industry standards, and creating closed-source ecosystems” under the name of safeguarding national security.

“These actions fragment global AI industry and supply chains, worsening the development imbalance in the global AI industry,” he wrote.

An editorial in state-backed tabloid Global Times over the weekend also dismissed the recent essay by Anthropic CEO Dario Amodei .

“On the surface, this appears to be a ‘rational statement’ focused on global AI security,” the editorial read. “A close reading, however, reveals its hidden agenda: The article is packed with containment provisions targeting China and is, in essence, a ‘Cold War playbook’ for the AI sector.”

While prominent executives and researchers in the US have gone public with their fears about the technology’s risks, raising similar concerns in China, where the ruling Communist Party tightly controls the media and surveils its population, is far more difficult and comes with much higher risks.

China’s technology sector also operates within a far more top-down regulatory environment, with the government setting extensive rules governing what AI systems can produce and how they are deployed. China’s national AI safety standards, for example, set requirements covering training data, privacy, security and prohibited content.

There is also greater public optimism about AI in China. A 2025 Edelman survey found that 72% of respondents in China said they trusted AI, compared with just 32% in the US.

And while Chinese AI labs have rapidly narrowed the gap with their American counterparts, they still trail the most advanced US systems. That may give Chinese companies and researchers less incentive to voluntarily constrain the development and deployment of increasingly capable AI systems as they seek to catch up.

There appears to be little reason for optimism. Despite a trade truce reached between the two countries in October last year, tensions have continued to flare.

The central obstacle is trust. Washington fears that cooperation could give China access to technology that strengthens its military and economic power, while Beijing sees “national security” as a US excuse to constrain China’s technological rise.

A post last month by Yuyuantantian , a social media account affiliated with Chinese state broadcaster CCTV, laid out the conditions for the US-China AI dialogue, saying Washington must guarantee its AI companies would follow any agreed-upon rules and regulations before any negotiations can take place.

“A clear distinction must be made between genuine safety threats and mere technological competition,” it said. “The real question is not ‘whether to govern,’ but who should define the safety boundaries of these frontier models.”

上周末，当全世界都在担忧不受约束的人工智能可能会毁灭人类时，美国和中国的领导人阐述了他们对这项划时代技术的愿景。

他们来自几乎截然相反的两端。

美国总统唐纳德·特朗普无视人工智能科学家以及 Anthropic、OpenAI 和谷歌等公司高管要求放缓人工智能发展的呼吁，称最重要的是在日益激烈的人工智能超级大国竞赛中保持美国对中国的领先地位。

“我们在人工智能领域领先于中国……坦白说，我希望保持这种领先地位，因为谁在人工智能领域获胜，谁就赢了，”他周六表示。

周日，中国国家主席习近平呼吁加强国际合作。他在印度向南方国家领导人发表讲话时表示，各国应“鼓励开源、开放、协作和共享”。他在讲话中重申了建立全球人工智能治理协议的提议，但并未详细说明该框架的具体内容。

特朗普和习近平的言论形成鲜明对比，凸显了这两个超级大国在争夺人工智能主导权的激烈竞争中缺乏共识和严重的信任赤字。

美国似乎越来越倾向于将自身在该领域的统治地位武器化，将其视为需要保护并出口到友好国家的战略资产，而落后于该领域的中国则提出了一种更加开放的模式。

人工智能将在两周后于华盛顿举行的备受瞩目的特朗普-习近平峰会上占据重要地位。在5月份于北京举行的会晤中，两人讨论了人工智能的监管措施，但成果有限。上周，中国商务部表示，两国领导人已同意建立人工智能政府间对话机制。但鉴于双方更广泛的地缘政治竞争，专家们对此次峰会取得任何实质性突破并不抱太大期望。

多年来，中国人工智能技术一直落后于美国竞争对手。这部分是由于美国的技术管制限制了中国企业获取先进的人工智能处理器——而这些处理器对于训练日益复杂的人工智能模型至关重要。

但去年，中国人工智能初创公司 DeepSeek 推出的一款模型震惊了全球业界，该模型使用远少于美国领先系统的计算资源，却达到了美国领先系统的性能。

从那时起，中国的人工智能实验室迭代速度加快，阿里巴巴、中信人工智能、Minimax 以及最近的 Moonshot 都推出了功能类似的强大模型，挑战了 Anthropic 和 OpenAI 等美国前沿公司的统治地位。

然而，随着中国模型的进步，美国对中国模型的作弊指控也愈演愈烈。华盛顿以及OpenAI和Anthropic等公司都声称，中国公司通过使用美国前沿模型的输出结果来训练自己的模型，窃取了美国的知识产权，这一过程在业内被称为“蒸馏”。

上周，包括联邦调查局、国家安全局和网络安全与基础设施安全局在内的美国联邦机构声称，中国人工智能公司正在对美国同行进行“工业规模”的窃取。

凯莉·库珀/路透社

上周，人道研究所发布了一份长篇报告，详细描述了包括据称与中国政府有关联的运营商在内的不良行为者企图滥用其系统来开发武器，并对新疆的维吾尔族等少数民族进行监视。

报告举例称，Anthropic公司发现，一名据信与中国军方有关联的用户向Moonshot公司的Kimi模型发送的查询被重定向到Claude模型，从而揭露了Moonshot公司秘密将用户请求转移到Claude模型的做法。报告还指出，其他中国公司，例如DeepSeek，也参与了类似的操作。

北京方面否认了这些指控，而中国的人工智能实验室大多保持沉默。上周，Moonshot拒绝就CNN的报道置评，而DeepSeek则未回应置评请求。

外交部发言人郭家坤周一表示：“散布各种威胁性言论，追求对抗和恶意竞争，只会扰乱全球人工智能治理的努力，对任何人都没有好处。”

他补充说：“各方应共同努力，促进开放、包容、易用且有益于人类的人工智能发展。”

正如美国越来越感到受到中国快速发展的模型的威胁一样，北京也对美国公司开发的系统的网络能力感到担忧——今年夏天，OpenAI 和 Anthropic 的代理程序失控并未经授权入侵其他系统，就展现了这些能力。

北京对人工智能安全风险的看法也与美国不同，它不仅关注人工智能失控等生存风险，还将美国视为对其自身技术发展的“霸权威胁”。除了优先考虑政治控制和社会稳定外，北京还把美国视为自身技术发展的“霸权威胁”。

国家安全部长陈义新在周日发表的一篇文章中，含蓄地批评美国以维护国家安全为名，“建立实体清单、实施技术管制、垄断行业标准、创建闭源生态系统”，强调了“霸权主义做法”的风险。

他写道：“这些举措使全球人工智能产业和供应链支离破碎，加剧了全球人工智能产业发展不平衡的局面。”

周末，官方小报《环球时报》的一篇社论也驳斥了人道协会首席执行官达里奥·阿莫迪最近发表的文章。

社论指出：“表面上看，这似乎是一份聚焦全球人工智能安全的‘理性声明’。然而，仔细阅读后就会发现其隐藏的意图：文章中充斥着针对中国的遏制条款，本质上是人工智能领域的‘冷战剧本’。”

虽然美国一些知名高管和研究人员已经公开表达了他们对这项技术风险的担忧，但在中国，由于执政的共产党严格控制媒体并监视民众，提出类似的担忧要困难得多，而且风险也高得多。

中国的科技行业也运行在一个更为自上而下的监管环境中，政府制定了广泛的规则来规范人工智能系统可以生成什么内容以及如何部署。例如，中国的国家人工智能安全标准就对训练数据、隐私、安全和禁止内容等方面提出了要求。

中国公众对人工智能的乐观程度也更高。爱德曼2025年的一项调查发现，72%的中国受访者表示他们信任人工智能，而美国只有32%的受访者持相同观点。

尽管中国的人工智能实验室已迅速缩小与美国同行的差距，但仍落后于美国最先进的系统。这或许会降低中国企业和研究人员在追赶美国的同时，主动限制日益强大的人工智能系统研发和部署的动力。

似乎没有什么理由保持乐观。尽管两国去年10月达成了贸易休战协议，但紧张局势依然持续升级。

核心障碍在于信任。华盛顿担心合作可能使中国获得增强其军事和经济实力的技术，而北京则认为“国家安全”是美国限制中国科技崛起的借口。

上个月，隶属于中国中央电视台的社交媒体账号“宇源探天”发布了一篇帖子，阐述了中美人工智能对话的条件，称华盛顿必须保证其人工智能公司遵守任何商定的规则和条例，谈判才能进行。

报告指出：“必须明确区分真正的安全威胁和单纯的技术竞争。真正的问题不在于‘是否应该监管’，而在于谁应该界定这些前沿模型的安全边界。”