中国的人工智能企业是否在暗中利用美国模型来缩小技术差距？在华盛顿及美国人工智能实验室Anthropic接连发出一系列指控后，这一疑问便成为中国科技行业关注的焦点。
Did Chinese AI companies secretly tap US models to close the technology gap? That is the central question hanging over China’s tech sector following a wave of allegations from Washington and American AI lab Anthropic.
根据美国国家安全局、联邦调查局以及网络安全与基础设施安全局上月联合发布的警告，中国的人工智能开发者一直在系统性地借鉴美国前沿模型的技术优势，以此降低自身在算力、电力消耗及研发方面的成本。
According to a joint advisory by the National Security Agency, FBI, and Cybersecurity and Infrastructure Security Agency last month, Chinese AI developers have been systematically extracting capabilities from US frontier models to reduce their own computing, electricity and research costs.
美国财政部长斯科特·贝森特也在7月份发出类似警告，称中国企业正在大规模实施所谓的“模型蒸馏”技术。
US Treasury Secretary Scott Bessent similarly warned in July that Chinese firms were engaged in large-scale “distillation”.
具体而言，Anthropic指控中国AI开发者利用虚假账户及代理网络窃取其旗舰模型Claude的推理数据；在某些情况下，这些企业甚至将该美国模型作为自家应用程序的幕后后端系统加以利用。
Specifically, Anthropic claimed Chinese AI developers used fraudulent accounts and proxy networks to siphon reasoning data from its flagship model, Claude – and in some cases even used the US model as a covert back end for their own apps.
下文将详细探讨这些指控是否站得住脚，以及为何中国政府对这些说法予以高度重视。Anthropic的指控内容是什么？
Here’s a closer look at whether the accusations hold water – and why Beijing is taking those claims seriously. What are Anthropic’s allegations?
Anthropic指责中国开发者实施了“工业级规模”的模型蒸馏——即借助性能更优越的“教师模型”的输出结果来优化另一款模型的技术性能。
Anthropic accused Chinese developers of conducting “industrial-scale” distillation, a technique in which outputs from a more capable “teacher” model are used to improve another model.
虽然模型蒸馏在整个AI行业都属于常见做法，但Anthropic认为中国企业此举已逾越了道德底线。据该公司上月发布的报告，中国相关实验室通过注册虚假账户并使用第三方代理服务，大规模获取Claude模型的输出内容及推理过程数据。
While model distillation is a common practice across the AI industry, Anthropic claimed Chinese firms crossed ethical lines. According to a report published by the company last month, Chinese labs circumvented usage restrictions through fraudulent accounts and third-party proxy services to harvest Claude’s output and reasoning traces at scale.
Anthropic指出，仅在10天时间里，Moonshot公司便向Claude Opus模型发送了近30万条用户请求。据称，这家中国企业借助由5380个虚假账户构成的代理网络来完成上述操作，这些账户大多注册于新加坡和日本。
Anthropic reported that over just 10 days, Moonshot routed almost 300,000 user requests, primarily to the Opus model. The Chinese company utilised a proxy network of 5,380 fraudulent accounts largely based in Singapore and Japan, Anthropic claimed.
此外，Anthropic声称一些公司将Claude用作隐秘的后端服务。该公司指出，这意味着用户以为自己正在与国产中文模型互动，但实际上他们的请求是由这家美国公司的服务器处理的。Moonshot方面未对置评请求作出回应。证据的可靠性如何？
Furthermore, Anthropic alleged that some companies used Claude as a hidden back end. This meant users believed they were interacting with a domestic Chinese model when their queries were actually being processed by the US firm’s servers, the firm said. Moonshot did not respond to a request for comment. How strong is the evidence?
业界分析师对于Anthropic调查结果的可信度及其归因存在分歧。
Industry analysts are divided over the credibility and attribution of Anthropic’s findings.
总部位于华盛顿的技术与国家战略研究中心联合创始人康斯坦丁·皮尔茨认为这些指控可信，并指出这反映了中国开发者在美国芯片出口限制下面临的巨大压力。
Konstantin Pilz, co-founder and fellow at the Washington-based Centre for Technology and Statecraft, considered the claims credible, noting that they reflected the intense pressure on Chinese developers facing US chip export restrictions.
“为了避免在人工智能模型性能方面进一步落后，中国AI企业必须大规模运用对抗性蒸馏技术，”皮尔茨说道。他同时指出，这种技术能帮助国内开发者省去生成高成本推理数据的环节。
“To not fall further behind on AI model performance, Chinese AI companies have to use adversarial distillation on a large scale,” Pilz said, suggesting that distillation allowed domestic developers to leapfrog expensive reasoning data generation.
美国新安全中心（CNAS）的研究助理本·海尤姆则表示，Anthropic长期以来一直能够收集相关威胁情报，从而识别并判定蒸馏攻击的来源。
Ben Hayum, a research assistant at the Centre for a New American Security (CNAS), said Anthropic had long been able to gather threat-intelligence signals that could detect and attribute distillation attacks.
不过，也有专家提出，要直接认定中国AI企业负有责任并非易事。
However, other experts argued that pinpointing direct blame on Chinese AI companies was not always straightforward.
布鲁金斯学会研究员凯尔·陈指出，第三方代理服务——即所谓的“中转站”——或许才是真正的罪魁祸首。他解释称，这些代理服务可能在未经中国AI实验室直接参与的情况下，暗中将用户请求转至Claude以获取并出售推理数据。
Kyle Chan, a fellow at the Brookings Institution, noted that third-party proxy services – or “transfer stations” – might be the real culprits. These proxies might secretly divert user requests to Claude to obtain and sell reasoning data without the direct involvement of the Chinese AI labs, he said.
深圳私募股权投资机构JG Investment专注于大语言模型与具身智能领域的投资经理王泽斌则认为，Anthropic的报告“并未提供足够的具体案例来独立验证相关企业究竟是如何获取及使用模型输出结果的”，他还指出报告中描述的具体蒸馏方法也值得商榷。为何北京方面会对此感到担忧呢？
Wang Zebin, an investment manager at Shenzhen-based private equity firm JG Investment focused on large language models and embodied intelligence, said Anthropic’s report “does not provide enough concrete examples to independently verify how those companies obtained or used the model outputs”, adding that the specific distillation methods described in the report were questionable. Why might Beijing be concerned?
据报道，在Anthropic发布相关报告后，中国监管机构开始对国内多家人工智能实验室展开调查。据《The Information》上月报道，国家互联网信息办公室曾询问了Anthropic点名的七家开发者，随后将调查重点转向深度求索（DeepSeek）与月之暗面（Moonshot）。深度求索方面未对置评请求作出回应。
Chinese regulators were reportedly investigating domestic AI labs following Anthropic’s report. According to a report by The Information last month, the Cyberspace Administration of China questioned seven developers named by Anthropic before narrowing its focus to DeepSeek and Moonshot. DeepSeek did not respond to a request for comment.
据Pilz分析，若该报道属实，北京方面最担忧的并非知识产权盗窃或模型抄袭，而是数据安全问题。
If the report was true, the main concern for Beijing would not be intellectual property theft or model copying, but data security, according to Pilz.
Anthropic称，用户发送给Claude的部分提示词中包含来自政府机构及企业的敏感信息，还包括姓名、电子邮箱及其他企业数据。
Anthropic claimed that some of the prompts sent to Claude contained sensitive information from state-affiliated organisations and companies, as well as names, email addresses, and other corporate data.
陈世荣指出，将此类敏感数据转交给美国的人工智能模型供应商可能给中国带来“极为严重的安保隐患”，其危害程度甚至超过搜索记录泄露——毕竟用户向AI系统提供的各类数据量极为庞大。尤其是编程相关的请求，极易导致底层IT基础设施信息或访问凭证外泄。
Re-routing sensitive data to American AI model providers may pose a “very significant security issue for China”, Chan said, adding that it would be “even more severe” than exposing search histories because users share so much data with AI. Coding requests in particular can leak underlying IT infrastructure or credentials.
Hayum表示：“中国理应对此予以高度重视。”他同时强调，监管机构必须审慎评估国内科技企业是否有能力保护敏感数据并有效应对新出现的网络安全风险。
“China should be seriously questioning this,” Hayum said, noting that regulators must evaluate whether they can trust domestic tech firms to safeguard sensitive data and manage emerging cybersecurity risks.
不过即便这些指控确有其事，陈世荣预计中国政府仍会低调处理此事，以免公开承认Anthropic的说法属实。
Even if the allegations were true, however, Chan expected Beijing to handle the matter discreetly to avoid publicly validating Anthropic’s claims.