If AI really could kill us all, why give any ground to China?

Go Nakamura/File/Reuters

China isn’t messing around.

The government recently unveiled a comprehensive strategy to make China the world’s AI leader by the end of this decade — a roadmap that includes massive, coordinated funding and resources on a scale that one-party rule makes significantly easier. Sure, the Chinese Communist Party is prone to hyperbole, but the country’s leaders view AI as an existential threat to their power , and they’re throwing the kitchen sink at the problem. They’re already catching up to the United States.

Meanwhile, the American AI industry says it wants to tap the brakes on its own development. Their fear is a different one: Silicon Valley is building a technology that could outstrip its makers’ control and eventually destroy humanity.

The likelihood and capability of AI’s destructive power is hotly debated and inconclusive. But the United States and China are nevertheless locked in an arms race that mirrors the Cold War’s nuclear standoff — the most dire of threats that forced the bitterest of enemies to come to the table and establish rules to protect humanity, even as both sides possessed enough power to blow up the world multiple times over.

Both the United States and China are feverishly debating the correct balance of AI regulation and technological advancement, and each side has distinct advantages in the race. Meanwhile, the global industry appears to be nearing a crucial, if ill-defined tipping point that could force developers and political leaders to reckon with AI’s power.

Alexander Kazakov/Sputnik/Reuters

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

America’s early AI lead has eroded.

The US government did what it could to maintain an advantage, restricting China from accessing top-of-the-line AI chips. That created serious obstacles for the country’s AI models.

But China’s AI companies developed a remarkable strategy to overcome that hurdle.

“They couldn’t catch up with Americans because they didn’t have the most advanced chips, so they made their software better and more power efficient,” said David Bellman, product lead at SynMax, a data intelligence company.

In February 2025, China’s DeepSeek AI model trailed OpenAI’s top model by just 0.4% in benchmark testing, a gigantic leap forward from previous models. Since then, Chinese and US models have been neck and neck, according to Stanford University’s 2026 AI Index report , which argued the US-China AI model performance gap is “effectively closed.”

Long Wei/VCG/Getty Images

The United States continues to hold major advantages in frontier-model production, private investment and data center infrastructure, noted Andrew Chin, an AI law professor at the University of North Carolina. So which country is “ahead” depends heavily on how you’re measuring.

But China is ceding nothing. In 2025, it unveiled its New Generation Artificial Intelligence Development Plan, a pathway to AI leadership that leverages the country’s party-run tech industry and government agencies to build out its infrastructure and train large language models. Among its goals: supercharge the Chinese economy to make it the world leader in transformative technology, including batteries, robots and automation.

“Central planning has allowed China to rapidly develop its sovereign technology and expand its global influence,” said University of Virginia National Security Data and Policy Institute researchers in a white paper published shortly after China introduced its policy.

But China’s strategy also contains some distinct disadvantages: Its AI invites political favoritism and encourages bias in its outputs, noted the UVA researchers. That’s because the country’s main fear resides not in the extinction of humanity but the threat to the Communist Party’s power. Its regulation of AI is mostly ideological, focusing mainly on the restriction of what the AI shares with its users and expanding the political control of the party.

That fear is evident in an article published Sunday by China’s State Security Minister Chen Yixin in the state-run China Cyberspace magazine. Chen warned that the country must maintain dominance in and control of AI, cautioning that the party could lose power over its people if it fails. The United States could use AI to gain a cybersecurity or espionage advantage over China, or the technology could be used to launch a propaganda war against the party, Chen argued.

Meanwhile, the American AI industry is at least giving lip service to self-regulation. AI leaders are meeting with politicians about reining in the technology, but President Donald Trump has said such efforts are conceding to hoaxes and conspiracy theories — and will offer a major concession to China.

It’s unclear how steadfast the industry will remain in its stated efforts to slow down its development or how much it believes its technology can really cause the human race harm. Skeptics note the American AI industry faces steep hurdles to gaining sufficient computing abilities to give its technology the power needed for truly destructive behavior. America’s data center buildout effort lags far behind the industry’s ambitious goals to increase its computing power. Shortages in labor, building supplies, high-end chips and permits, in addition to rising interest rates and public antipathy to data centers, have caused wide-scale data center construction delays.

Brandon Bell/Getty Images

“Those are fundamental problems that could torpedo tech business models,” noted Darrell West, senior fellow of governance studies at the Brookings Institution.

But the fact of the matter is: No one really knows how much computing power is needed for AI to kill everyone.

“Anyone who tells you that once we reach such and such capacity, disaster happens, and up to that point we’re safe, is lying,” said Vincent Conitzer, computer science professor at Carnegie Mellon. “For all we know, with some clever ideas added, today’s capacity is already enough.”

That’s why advocates for AI regulation believe the efforts to rein in the technology must begin irrespective of the industry’s anticipated X date for reaching a level of artificial intelligence that could be destructive.

“Infrastructure delays may give us breathing room, but breathing room is not governance,” said Chin, the AI law professor.

Anthropic CEO Dario Amodei argued in an open letter this weekend that the United States must maintain dominance over China in AI capabilities, but the pace of US innovation shouldn’t exceed what’s needed to outpace its rival. That will give the American industry time to ensure AI isn’t dangerous while creating leverage over China to force it to the negotiating table to form some kind of global regulation.

“If we slow down by more than this amount, then (unpaced) Chinese Communist Party-associated projects will pull ahead, creating significant national security risk,” Amodei said. “Some may believe these measures make it more difficult to cooperate with China, but I believe the opposite is true: these measures increase the leverage held by democracies and make an agreement more likely in the future.”

China appears unenthused by Amodei’s proposal. An editorial in state-owned tabloid Global Times called the open letter a “Cold War playbook.”

David Paul Morris/Bloomberg/Getty Images

Nevertheless, China, perhaps even more than the United States, also fears out-of-control AI that could act outside of the government’s control. It, too, recognizes the dire need to rein in the technology to ensure it acts within its bounds.

“Beijing likely has its own deep concerns about AI’s potential to destabilize its social and political system,” said Philip Potter, executive director of the University of Virginia’s National Security Data and Policy Institute.

That’s why a spirit of cooperation may be in the best interests of both countries’ governments and industries, and arrangements where both countries restrain private-sector AI development while walling off AI development as a state and national security enterprise could be something both countries’ governments and private industries agree to, Potter said.

The alternative is a self-justifying arms race, Chin argues: If each country believes it must maintain a strategic advantage, then no one can have a safety policy.

“There is certainly a legitimate national-security interest in preventing a strategic competitor from obtaining a decisive technological advantage,” Chin said. “But there is also a national-security interest in not creating systems whose behavior, objectives or downstream effects we cannot reliably understand or control.”

Go Nakamura/档案/路透社

中国可不是闹着玩的。

中国政府近期公布了一项全面战略，旨在到本十年末使中国成为全球人工智能领域的领导者——这项路线图包括大规模、协调一致的资金和资源投入，而一党专政使得这一规模的投入更容易实现。诚然，中国共产党有时会夸大其词，但中国领导人确实将人工智能视为对其政权的生存威胁，并为此不遗余力。他们已经在追赶美国。

与此同时，美国人工智能产业表示希望放缓自身发展步伐。他们的担忧有所不同：硅谷正在研发的技术可能会超出其创造者的控制，并最终毁灭人类。

人工智能破坏力的可能性和能力一直备受争议，尚未有定论。但美国和中国却陷入了一场军备竞赛，这场竞赛与冷战时期的核僵局如出一辙——正是这种最严峻的威胁迫使昔日最敌对的双方坐到谈判桌前，制定规则以保护人类，尽管当时双方都拥有足以毁灭世界数次的强大力量。

美国和中国都在热烈讨论人工智能监管与技术进步之间的平衡之道，双方在这场竞赛中都拥有明显的优势。与此同时，全球人工智能产业似乎正接近一个关键的、尽管定义尚不明确的转折点，这可能会迫使开发者和政治领导人正视人工智能的力量。

亚历山大·卡扎科夫/人造卫星/路透社

特朗普不会因为想打败中国而限制人工智能的发展。北京究竟有多大的威胁？

美国在人工智能领域的早期领先优势已经消失。

美国政府竭尽所能保持优势，限制中国获取最先进的人工智能芯片。这给中国的人工智能模型发展造成了严重障碍。

但中国的人工智能公司制定了一项非凡的战略来克服这一障碍。

“他们无法赶上美国人，因为他们没有最先进的芯片，所以他们改进了软件，提高了软件的能效，”数据智能公司 SynMax 的产品负责人 David Bellman 说。

2025年2月，中国的DeepSeek人工智能模型在基准测试中仅以0.4%的差距落后于OpenAI的顶级模型，这相比之前的模型而言是一次巨大的飞跃。斯坦福大学2026年人工智能指数报告指出，此后中美模型的性能一直不相上下，该报告认为中美人工智能模型性能差距已“基本消除”。

龙伟/VCG/Getty Images

北卡罗来纳大学人工智能法学教授安德鲁·钦指出，美国在前沿科技生产、私人投资和数据中心基础设施方面仍然拥有显著优势。因此，哪个国家“领先”很大程度上取决于衡量标准。

但中国并未做出任何让步。2025年，中国发布了《新一代人工智能发展规划》，这是一条通往人工智能领导地位的道路，它将利用中国党控科技产业和政府机构来建设基础设施并训练大型语言模型。其目标之一是：大力推动中国经济发展，使其成为包括电池、机器人和自动化在内的变革性技术领域的全球领导者。

“中央计划经济使中国能够迅速发展其自主技术并扩大其全球影响力，”弗吉尼亚大学国家安全数据与政策研究所的研究人员在中国推出该政策后不久发表的一份白皮书中指出。

但弗吉尼亚大学的研究人员指出，中国的人工智能战略也存在一些明显的弊端：它容易滋生政治偏袒，并导致输出结果带有偏见。这是因为中国的主要担忧并非人类灭绝，而是对共产党政权的威胁。中国对人工智能的监管主要基于意识形态，重点在于限制人工智能与用户共享的信息，并扩大党的政治控制。

这种担忧在中国国家安全部长陈宜新周日发表于官方媒体《中国网络空间》杂志的一篇文章中显露无疑。陈宜新警告说，中国必须保持对人工智能的主导地位和控制力，并告诫说，如果失败，党可能会失去对人民的领导。陈宜新认为，美国可能利用人工智能在网络安全或间谍活动中对中国取得优势，或者这项技术可能被用来对党发动宣传战。

与此同时，美国人工智能产业至少在口头上支持自我监管。人工智能领域的领导者们正与政界人士会面，商讨如何限制这项技术的发展，但唐纳德·特朗普总统却表示，这些努力是在向骗局和阴谋论妥协，并将向中国做出重大让步。

目前尚不清楚人工智能行业在减缓自身发展速度方面究竟能坚持多久，也不清楚他们究竟认为这项技术会对人类造成多大的危害。怀疑论者指出，美国人工智能行业在获得足够的计算能力方面面临着重重阻碍，而这些能力正是其技术发挥真正破坏性作用所必需的。美国的数据中心建设进度远远落后于该行业提升计算能力的宏伟目标。劳动力、建筑材料、高端芯片和许可证的短缺，加上利率上升和公众对数据中心的抵触情绪，都导致了数据中心建设的大规模延误。

Brandon Bell/Getty Images

“这些都是可能摧毁科技商业模式的根本性问题，”布鲁金斯学会治理研究高级研究员达雷尔·韦斯特指出。

但事实是：没有人真正知道人工智能需要多少计算能力才能杀死所有人。

卡内基梅隆大学计算机科学教授文森特·科尼策尔说：“任何告诉你一旦我们达到某个特定的能力水平，灾难就会发生，而在此之前我们都是安全的，都是在撒谎。就我们所知，只要加入一些巧妙的想法，今天的能力可能就已经足够了。”

正因如此，人工智能监管的倡导者认为，无论业界预期人工智能何时达到可能具有破坏性的水平，都必须开始努力控制这项技术。

“基础设施建设的延误或许能给我们一些喘息之机，但喘息之机并非治理之道，”人工智能法学教授秦说道。

Anthropic首席执行官达里奥·阿莫迪本周末在一封公开信中指出，美国必须在人工智能领域保持对中国的领先地位，但美国的创新速度不应超过超越竞争对手所需的速度。这将使美国业界有时间确保人工智能不会带来危险，同时也能增强对中国的影响力，迫使其回到谈判桌前制定某种全球监管法规。

阿莫迪表示：“如果我们放缓的步伐超过这个幅度，那么（不受控制的）与中国共产党相关的项目将会加速推进，从而造成重大的国家安全风险。有些人可能认为这些措施会使与中国的合作更加困难，但我认为事实恰恰相反：这些措施增强了民主国家的谈判筹码，并增加了未来达成协议的可能性。”

中国似乎对阿莫迪的提议并不热衷。《环球时报》发表社论，称这封公开信是“冷战剧本”。

David Paul Morris/Bloomberg/Getty Images

然而，中国或许比美国更甚，同样担忧人工智能失控，可能不受政府控制。中国也认识到迫切需要对这项技术进行管控，确保其在既定范围内运行。

“北京方面可能对人工智能可能破坏其社会和政治体系的稳定性深感担忧，”弗吉尼亚大学国家安全数据与政策研究所执行主任菲利普·波特表示。

波特表示，正因如此，合作精神可能符合两国政府和产业的最佳利益，两国限制私营部门人工智能发展，同时将人工智能发展划为国家和国家安全事业的安排，可能是两国政府和私营企业都能接受的。

秦认为，另一种选择是陷入自我辩解的军备竞赛：如果每个国家都认为自己必须保持战略优势，那么就没有人能够制定安全政策。

“防止战略竞争对手获得决定性的技术优势，当然符合国家安全利益，”秦说。“但同样符合国家安全利益的是，不应建立我们无法可靠地理解或控制其行为、目标或后续影响的系统。”