DAVID PIERSON, ADAM SATARIANO2026年10月22日根据上周提出的规则，下一代先进人工智能模型出口到中国境外，可能需要获得政府的许可。 Aly Song/ReutersWhen the Chinese start-up Moonshot AI in July released Kimi K3, an artificial intelligence model that rivaled some of the best systems made by American companies, it was free for anyone to download.今年7月，当中国初创企业月之暗面发布人工智能模型Kimi K3时，该模型可供任何人免费下载；其性能已足以与美国公司开发的高水平系统相媲美。By making advanced A.I. models easily available, Chinese companies have turned open access into one of China’s most powerful advantages in its rivalry with the United States to shape the future of the technology.通过推动先进人工智能模型的便捷普及，中国企业已将开放获取打造为中国与美国争夺该技术未来主导权时的核心竞争优势之一。Now Beijing is moving to tighten its grip on the global spread of the country’s top-tier A.I., proposing rules that could give the government new power to track, and potentially limit, where the systems go.如今，北京正采取措施加强对本国顶尖人工智能全球传播的管控，拟议中的新规或将赋予政府跟踪乃至限制这些系统去向的新权力。At the same time, the government sought to reassure users that the move would not stifle access to Chinese A.I., saying the vast majority of the technology would still be freely exported.与此同时，政府也试图向用户澄清，此举并非要阻碍对中国人工智能的使用，并表示绝大多数技术仍可自由出口。The draft rules reflect Beijing’s effort to balance two competing goals: regulate the spread of its frontier A.I. systems for strategic and security reasons while still encouraging global users to adopt Chinese software.规则草案折射出北京在平衡两大相互竞争的目标：一方面出于战略与安全考量监管前沿人工智能系统的传播，另一方面仍鼓励全球用户采用中国软件。“It is almost like you have two voices speaking at the same time,” said Kyle Chan, a fellow at the Brookings Institution focusing on Chinese technology and industrial policy. “One saying, ‘We want extra control of these sensitive technologies so that they can’t hurt China’ and another voice saying, ‘We want to develop the best A.I. technology and we want to share it with rest of the world.’”布鲁金斯学会关注中国科技与产业政策的研究员陈凯欣(Kyle Chan)表示：“这几乎就像有两个声音在同时发声：一个声音说‘我们需要对这些敏感技术实施额外管控，以免危害中国’；另一个声音则说‘我们要开发最顶尖的人工智能技术，并与世界共享’。”今年7月在上海举行的人工智能大会上，中国人工智能初创公司月之暗面为其模型Kimi设立的展台。Kimi K3可与美国公司打造的一些最佳系统相媲美。China’s Ministry of Commerce’s announcement of the draft rules last Friday threw uncertainty into an A.I. market that has come to see the availability of free Chinese A.I. models from companies like Alibaba, DeepSeek, Z.ai and Moonshot as an important alternative to services from American companies such as Anthropic and OpenAI.上周五，中国商务部发布规则草案公告，给当前的人工智能市场蒙上了不确定性。此前，市场已逐渐将阿里巴巴、DeepSeek、智谱及月之暗面等公司提供的免费中国人工智能模型视为替代Anthropic和OpenAI等美国公司服务的重要选择。The ministry said the rules would not affect Chinese models that have already been released, and overseas users could still access even the most advanced systems through apps and online services.商务部表示，新规不会影响此前已发布的中国模型，海外用户未来仍可通过应用程序及在线服务使用包括最先进系统在内的各类模型。The Risks China Wants to Contain中国拟遏制的风险If passed, the rules would require Chinese companies to obtain licenses to export the core technology behind their leading models.一旦通过，这些规则将要求中国企业在出口其领先模型背后的核心技术时必须取得许可。They would also ban the export of A.I. that can be used to launch cyberattacks, as well as models trained on what China calls “core data.” The government broadly defines that as data that concerns national security, the welfare of Chinese citizens and the public interest. Experts say the scope includes data on the military or critical infrastructure, state secrets and other sensitive information.新规还将禁止出口可用于发起网络攻击的人工智能，以及基于中国所称“核心数据”训练出的模型。中国政府将“核心数据”广泛定义为涉及国家安全、中国公民福祉及社会公共利益的数据。专家指出，这一范围涵盖军事或关键基础设施、国家秘密及其他敏感信息相关数据。The commerce ministry said the new controls were necessary to guard against abuse and malicious use and to ensure that A.I. always remains under human control.商务部指出，出台新的管控措施对于防范技术滥用和恶用、确保人工智能始终处于人类掌控之下是必要的。The power of A.I. systems to escape their creators’ control and hack into real-world systems without human knowledge has drawn global scrutiny. In the United States, A.I. agents being tested by OpenAI broke into outside systems including the A.I. start-up Hugging Face, and meddled with U.S. government websites. The agents also hacked into an Australian government website and acquired health data, prompting an investigation by Australian officials.人工智能系统摆脱开发者控制并在人类不知情下入侵现实世界系统的能力已引发全球关注。在美国，OpenAI测试的AI智能体曾入侵外部系统（包括人工智能初创公司Hugging Face），并擅闯美国政府网站。此外，相关智能体还入侵了澳大利亚政府网站并获取了健康数据，引发澳大利亚官方调查。In China, the internet giant Tencent last week acknowledged that A.I. agents it used to evaluate other companies’ models had behaved in ways it did not intend, after researchers said the agents tried to bypass anti-robot protections on websites.在中国，研究人员指出互联网巨头腾讯用于评估其他公司模型的AI智能体试图绕过网站的反爬虫保护后，腾讯于上周承认，这些智能体的行为确实超出了设计初衷。China’s proposed export rules signaled that the government is concerned about the safety of increasingly sophisticated A.I. models, as calls have grown for the United States and China to prevent the technology from causing catastrophic harm. Earlier this month, the Communist Party’s Central Committee and the State Council, China’s cabinet, called for the establishment of a monitoring and emergency response system that would ensure that A.I. is “safe, reliable and controllable.”在中美两国防范人工智能引发灾难性危害的呼声日益高涨之际，中国拟议的出口规则表明官方对日益复杂的人工智能模型的安全性心存担忧。本月早些时候，中共中央、国务院要求建立监测与应急响应体系，确保人工智能“安全、可靠、可控”。“These controls show that China is taking the risks of powerful A.I. much more seriously than was generally understood,” said Martin Chorzempa, senior fellow at the Peterson Institute for International Economics.彼得森国际经济研究所高级研究员马丁·肖尔赞帕表示：“这些管控举措表明，中国对待强大人工智能风险的严肃程度远超外界普遍预期。”What the Rules Could Change新规可能带来的改变Still, much remains unclear, such as how strictly China will enforce the rules and how burdensome the licensing process will be, especially for labs that release updates to their models every few weeks.不过，目前仍有许多细节尚不明确，例如中国执行这些规则的严厉程度，以及相关许可流程的繁琐程度——特别是对于每隔数周即更新发布一次模型的实验室而言。One of the draft rules called for export restrictions on any “globally leading” model, which the ministry said could be partly defined as a system that ranks in the top 10 for global usage. It also set thresholds for a model’s size and how much computation was used in its training. Meeting any one of those criteria would be enough to require a license.规则草案提及，将对所有“全球领先”的模型实施出口限制。商务部表示，其衡量标准之一包括在全球使用量中排名前十的系统。草案还对模型的参数规模及训练计算量设定了门槛，满足上述任意一项标准即需申请许可。Analysts at Trivium China, a research and advisory firm, wrote in a note to clients that the criteria would capture future versions of Kimi, Qwen, and DeepSeek, referring to three of the most advanced Chinese models.策纬咨询(Trivium China)分析师在一份面向客户的报告中写道，这些标准将覆盖Kimi、通义千问和DeepSeek等三大中国顶尖模型的未来迭代版本。通义千问是中国科技巨头阿里巴巴于2023年启动的人工智能模型系列。阿里巴巴将其最受欢迎的人工智能模型开源，允许他人免费使用和修改。If adopted, the rules would also give Beijing a say over who buys Chinese A.I. companies.一旦通过，这些规则还将赋予北京对中国人工智能企业收购案的审批决策权。That could close the regulatory loophole that briefly allowed Meta to acquire Manus, a Singapore-based A.I. company that was founded in China, earlier this year. Beijing ultimately scuttled the deal, and in an apparent effort to discourage Chinese A.I. executives from moving businesses offshore, restricted Manus executives from leaving China.这可能堵住今年早些时候一度让Meta得以收购Manus的监管漏洞。Manus是一家在中国创立、总部位于新加坡的人工智能公司。北京最终阻止了这笔交易，并限制Manus高管离境，以此警告中国人工智能高管切勿将业务转移至海外。No Longer Fully Open?不再完全开放？Many Chinese models are open-weight, which means they are available for other developers to download and customize as they see fit. That contrasts with services like ChatGPT or Claude, which only offer access through a subscription and cannot be modified.许多中国模型属开放权重模型，这意味着其他开发者可以随意下载并根据自身需求进行定制。这与ChatGPT或Claude等仅限付费订阅且无法被修改的服务形成了鲜明对比。Once downloaded, a Chinese model could be run on a company’s own computers indefinitely, beyond the reach of anyone who might want to restrict it. Users of American A.I., on the other hand, are essentially renting access that can be taken away. That was the case in June when the Trump administration temporarily restricted foreign access to Anthropic’s Fable 5 model, citing national security risks.中国模型一旦下载，即可在企业自身的计算机上无限期运行，不受任何后续限制措施的影响。相比之下，美国人工智能的用户本质上只是租用使用权，这种权限随时可能被剥夺。例如今年6月，特朗普政府就曾以国家安全风险为由，临时限制外部获取Anthropic的Fable 5模型。Now companies outside China that rely on Chinese A.I. systems could face similar risks.如今，中国以外依赖中国人工智能系统的企业也可能面临类似的风险。“This could make U.S. businesses more hesitant to rely on Chinese models going forward,” said Chris McGuire, a senior fellow at the Council on Foreign Relations.美国外交关系协会高级研究员克里斯·麦奎尔表示：“这可能会使美国企业在未来对依赖中国模型持更加审慎的态度。”A.I. has emerged as a point of geopolitical tension between Washington and Beijing, with U.S. agencies accusing DeepSeek and other Chinese A.I. companies of industrial-scale copying. On Monday, the Department of Homeland Security ordered the removal of DeepSeek’s tools from the technological systems of many U.S. government agencies, which China described as an act of discrimination.人工智能已成为华盛顿与北京地缘政治摩擦的焦点之一，美国相关部门指责DeepSeek等中国人工智能企业进行工业级抄袭。周一，美国国土安全部下令在多个政府机构的技术系统中全面停用DeepSeek的工具，对此中国指责是歧视性做法。A.I. as a Tool of Influence人工智能作为软实力工具Beijing, however, has also couched the controls in the language of international cooperation, in a nod to the soft power influence of its A.I. technology.然而，北京也将相关管控置于国际合作的语境下，以此彰显其人工智能技术的软实力影响。北京的一处DeepSeek标志。DeepSeek的开源模型是中国实现成为人工智能超级大国雄心壮志的核心。The government pledged to help narrow the digital divide in the world, particularly by supporting developing countries that are members of the World Artificial Intelligence Cooperation Organization, a group that China founded in July that includes Indonesia, Pakistan and Brazil.中国政府承诺帮助缩小全球数字鸿沟，特别是加大对世界人工智能合作组织成员国中发展中国家的支持力度。该组织由中国于今年7月发起成立，成员包括印度尼西亚、巴基斯坦和巴西等国。“That’s potentially quite noteworthy,” said Julian Gewirtz, a former China director for the National Security Council in the Biden administration now a scholar at Columbia University. “Beijing is establishing controls while suggesting it would offer favorable treatment for cooperation through a China-backed institution.”曾任拜登政府国家安全委员会中国事务主管、现为哥伦比亚大学学者的朱利安·格维兹指出：“这在潜在层面相当值得关注。北京在确立管控的同时表明，它会通过一个中国支持的机制，为愿意合作的各方提供特惠待遇。”郭莉莉(Lily Kuo)和Berry Wang对本文有报道贡献。David Pierson报道中国外交政策和中国与世界的经济与文化交互。他从事新闻工作已超过20年。Adam Satariano是时报科技记者，常驻伦敦。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