KYLE CHAN2026年9月18日 Illustration by Sam Whitney/The New York TimesAnthropic’s chief executive, Dario Amodei, recently said the world should collectively slow the development of A.I. until we get a better handle on the safety risks this new technology poses. Sam Altman of OpenAI and Elon Musk voiced their support for this plan.Anthropic首席执行官达里奥·阿莫迪最近表示，世界应该共同放缓人工智能的发展，直到我们能更好地应对这项新技术带来的安全风险。OpenAI的萨姆·奥特曼和埃隆·马斯克对这一计划表达了支持。American A.I. companies know they can’t tap the brakes on A.I. by themselves — China has to go along. Chinese companies such as DeepSeek, Z.ai and Moonshot have A.I. models nearly as powerful as Anthropic’s and OpenAI’s. And they likely pose many of the same risks; terrorists could potentially use their tools to carry out cyberattacks or design biological weapons. Many people hope the upcoming summit between President Trump and China’s leader, Xi Jinping, could be an opportunity to make some headway on such a partnership.美国的人工智能公司知道，仅凭自身力量无法为人工智能发展踩下刹车——中国必须参与进来。深度求索、智谱AI和月之暗面等中国公司拥有的人工智能模型能几乎与Anthropic和OpenAI的模型一样强大。它们也很可能构成许多相同的风险，恐怖分子可能会利用它们的工具来实施网络攻击或设计生物武器。许多人希望，特朗普总统与中国领导人习近平即将举行的峰会能成为推动此类合作取得进展的契机。Too often, however, A.I. safety experts assume that China has the same concerns as we do about A.I. The reality is, the Chinese have a different set of worries. And until we understand what those are, cooperation with China will be impossible.然而，人工智能安全专家往往认为，中国对人工智能有与我们相同的担忧。现实情况是，中国人的担忧截然不同。除非我们了解这些担忧的具体内容，否则与中国合作将是不可能的。Generally speaking, China doesn’t fret as much about A.I. as an existential threat to humanity. You don’t see its political leaders or the founders of Chinese A.I. companies warning that superhuman A.I. is “probably the greatest threat to the continued existence of humanity,” as Mr. Altman wrote in 2015. It’s not central to Chinese A.I. policymaking, nor a part of Chinese tech culture.总的来说，中国并没有那么担心人工智能会成为人类的生存威胁。你不会看到中国政治领导人或中国人工智能公司的创始人警告说，超越人类的人工智能“可能是人类继续生存的最大威胁”，正如奥特曼在2015年所写的那样。这并非中国人工智能决策的核心，也不属于中国科技文化的一部分。That’s mainly because of differences in how the two countries think about A.I. training and development. The leading American A.I. companies are hyper-focused on advancing “recursive self-improvement” — using A.I. systems themselves to develop even better A.I. systems. There are fears that this feedback loop could accelerate the pace of A.I. development to the point where humans no longer have control over the process. American tech leaders believe this takeoff point might be around the corner.这主要是因为两国在如何看待人工智能训练和发展方面存在差异。美国领先的人工智能公司极度关注推进“递归自我改进”——即利用人工智能系统自身来开发更好的人工智能系统。人们担心，这种反馈循环可能会加速人工智能的发展步伐，最终导致人类无法再掌控这一进程。美国科技界领袖认为，这个起飞点可能就在眼前。Chinese A.I. companies are also working on recursive self-improvement. But because of their limited access to A.I. computing power and the slower pace of progress, they believe the takeoff moment is still technologically a long way off. Chinese policymakers seem to share that outlook and so are more focused on building resilient A.I. supply chains and increasing A.I. adoption throughout the economy.中国的人工智能公司也在致力于递归自我改进。但由于他们获取人工智能算力的途径有限，且进展步伐较慢，他们认为，从技术上讲，这个起飞时刻还有很长的路要走。中国决策者似乎也认同这一前景，因此他们更专注于建立有韧性的人工智能供应链，并提高人工智能在整个经济中的应用。So why would the Chinese agree to a slowdown in A.I. development if they don’t think a crisis is imminent? If anything, Chinese A.I. companies are trying to move faster, hoping to narrow the performance gap with American A.I. models.如果中国人不认为危机迫在眉睫，他们为什么要同意放缓人工智能的发展呢？事实上，中国的人工智能公司正试图加快步伐，希望缩小与美国人工智能模型之间的性能差距。For the moment, Chinese policymakers are more worried about the political and social risks caused by A.I. They don’t want it creating deepfakes of political leaders or saying the “wrong” things about Taiwan or Tiananmen Square. They don’t want A.I. chatbots to create psychological dependencies or threaten child safety, or set off social instability or a mass movement powerful enough to topple the Communist Party.目前，中国决策者更担心人工智能引发的政治和社会风险。他们不希望它制造政治领导人的深度伪造内容，或在台湾或天安门广场问题上说“错”话。他们不希望人工智能聊天机器人造成心理依赖、威胁儿童安全，或者引发社会动荡，抑或掀起一场强大到足以推翻共产党的群众运动。China’s A.I. regulations today are generally aimed at these types of risks. Most new public-facing A.I. models must be registered with the government and pass a battery of tests before they can be released. China has created extensive rules requiring labeling of A.I.-generated content and limiting how humanlike A.I. chatbots can operate.中国目前的人工智能法规大多针对这类风险。大多数面向公众的新人工智能模型必须在政府备案，并在发布前通过一系列测试。中国制定了广泛的规则，要求对人工智能生成的内容进行标识，并限制人工智能聊天机器人模拟人类行为的程度。A spokesman for China’s foreign ministry has dismissed U.S. calls for a slowdown as “fear-mongering” intended to hold the country back in the race for A.I. dominance. (Indeed, Mr. Trump has said he opposes a slowdown specifically to prevent China from catching up.)中国外交部的一位发言人驳斥了美国关于放缓发展的呼吁，称其为“危言耸听”，旨在阻碍中国在人工智能主导权竞争中的进展。（事实上，特朗普曾表示，他反对放缓发展，正是为了防止中国迎头赶上。）There are some, limited signs that Beijing is starting to take large-scale A.I. threats more seriously. Two types of risks have received special attention in China.有一些有限的迹象表明，北京正开始更加认真地对待大规模的人工智能威胁。有两类风险在中国受到了特别关注。The first is cybersecurity. This year, many Chinese users downloaded a popular A.I. agent tool called OpenClaw and used it to automate computer-based tasks. Chinese officials quickly issued a warning that A.I. tools like OpenClaw could expose users to data leaks and cybersecurity vulnerabilities.第一类是网络安全。今年，许多中国用户下载了一款名为OpenClaw的流行人工智能智能体工具，并用它来自动化处理计算机任务。中国官员迅速发出警告，称OpenClaw这类人工智能工具可能会让用户面临数据泄露和网络安全漏洞风险。Then in April, Anthropic announced that it had trained a new model called Mythos that demonstrated significantly elevated cyberattack abilities. It did not take long for the Chinese to appreciate the national security implications. The head of China’s main security agency recently issued a rare warning that the cybersecurity landscape was changing, and critical infrastructure was now vulnerable to A.I.-led cyberattacks from adversaries. He pointed to some of the latest models from OpenAI and Anthropic.今年4月，Anthropic宣布他们训练了一个名为Mythos的新模型，该模型表现出显著提升的网络攻击能力。中国人没过多久就意识到了其对国家安全的影响。中国主要安全机构的负责人最近发出了一次罕见的警告，指出网络安全格局正在发生变化，关键基础设施现在很容易受到对手发动的、由人工智能主导的网络攻击。他特别提到了OpenAI和Anthropic的一些最新模型。Adding to these concerns, a U.S.-based security company recently used A.I. to develop an experimental tool that could rapidly compromise accounts on WeChat, the Chinese messaging platform with 1.4 billion monthly users. Although the tool was not released into the wild, it gave the Chinese pause.加剧这些担忧的是，一家总部位于美国的安全公司最近利用人工智能开发了一款实验性工具，该工具可以迅速攻破微信上的账户，微信是一个拥有14亿月活用户的中国即时通讯平台。尽管该工具并未公开发布流传，但它还是引起了中方的警觉。The second A.I. risk that’s gaining attention in China is runaway A.I operating beyond human control. China has watched closely as rogue American models have broken out and breached other websites’ security. The OpenAI models’ attack in July on Hugging Face, an A.I. infrastructure company, has been widely covered in Chinese state media.在中国引起关注的第二类人工智能风险是失控的人工智能超出人类控制的范围运作。中国一直在密切关注失控的美国模型突破限制并攻破其他网站安全防线的事件。今年7月OpenAI模型对人工智能基础设施公司Hugging Face的攻击，已在中国官方媒体上被广泛报道。Runaway A.I. is being talked about in higher levels of China’s government. Mr. Xi mentioned the risks of “loss of control” in July. More recently, a senior cybersecurity official in its Cyberspace Administration warned of “extreme loss of control,” citing the A.I. breakout incidents in the United States.失控的人工智能正成为中国政府高层讨论的话题。习近平在7月提到了“失控”的风险。最近，中国国家互联网信息办公室的一名高级网络安全官员援引美国的人工智能失控事件，警告称存在“极端失控”的危险。These shifting views of A.I. risks suggests both countries may be motivated to cooperate on A.I. safety out of sheer national interest. Neither country wants terrorists to use A.I. to create viruses. Neither country wants autonomous A.I. systems to accidentally hack its companies or institutions.这些对人工智能风险不断转变的看法表明，两国出于国家利益的考量，可能有动力在人工智能安全方面开展合作。两国都不希望恐怖分子利用人工智能制造病毒。两国也不希望自主的人工智能系统意外黑入本国的公司或机构。On the other hand, each country’s suspicions of the other will make any kind of agreements about A.I. safety difficult. China is very worried that America could use A.I. to enhance its cyber and military weapons aimed at China. And same is true in the other direction. Without a base line level of mutual understanding, the world’s two A.I. superpowers are bound to talk past each other.另一方面，两国对彼此的猜疑将使在人工智能安全方面达成任何形式的协议变得困难。中国非常担心美国会利用人工智能来增强针对中国的网络和军事武器。反之亦然。如果没有最基本的相互理解，世界上这两个人工智能超级大国注定只会自说自话。Fortunately, that common ground is possible. Rather than jump straight into debates over a global slowdown or an arms control treaty for A.I., the United States and China should start with something much more basic: realizing and acknowledging how the other side sees A.I. risks. From there, they can start to explore areas where both sides share an interest in mitigating threats from third parties or even the technology itself.幸运的是，这种共识是有可能达成的。中美两国与其直接跳入关于全球放缓或人工智能军备控制条约的争论，不如从更基本的事情开始：理解并承认对方如何看待人工智能风险。在此基础上，他们可以开始探索双方在减轻来自第三方甚至技术本身威胁方面具有共同利益的领域。There will be disagreements to resolve, and plenty of quibbles on details to sort out. But if the two countries can agree they want to avoid A.I. catastrophe, cooperation on A.I. safety is not such an unrealistic possibility.虽然会有分歧需要解决，在细节上也会有诸多争论需要理清。但如果两国能够达成一致，确认双方都希望避免人工智能灾难，那么在人工智能安全领域开展合作并非遥不可及。Kyle Chan是布鲁金斯学会的研究员，他是中国技术发展和产业政策方面的专家，也是行业通讯"High Capacity"的作者。图片来源：Getty Images翻译：纽约时报中文网点击查看本文英文版。获取更多RSS：https://feedx.net https://feedx.site