A Chinese company has been trying to develop artificial intelligence-powered technology that would enable authoritarian governments to not just monitor dissidents but also potentially predict who could become one in the future.

一家中国公司一直在尝试开发一项由人工智能驱动的技术，不仅能让威权政府监控异见人士，还能潜在地预测未来谁可能成为异见人士。

The work, which appears to be in the research stage, is ripped out of dystopian science fiction, offering a glimpse of a world in which an authoritarian state is able to move against its citizens before they begin any public dissent.

这项似乎还处于研究阶段的工作简直如同出自反乌托邦科幻小说，让我们得以一瞥一个威权国家在公民尚未公开表达异议之前就对其采取行动的世界。

The Chinese company, Geedge Networks, sells a commercial version of the Great Firewall, the surveillance and censorship software that China uses to control online activity. Those tools allow governments to monitor internet traffic and flag when someone tries to get around traditional internet censorship.

这家名为积至的中国公司销售商业版的“防火长城”——中国用于控制网络活动的监控和审查软件。这些工具使政府可以监控互联网流量，并在有人试图绕过传统互联网审查时发出警告。

But according to leaked company documents, the firm is working on new products that use artificial intelligence to examine location data and internet use to predict who could do or say something critical of the government, according to researchers at Vanderbilt University.

但范德比尔特大学的研究人员表示，根据泄露的公司文件，该公司正在开发使用人工智能审查位置数据和互联网使用情况的新产品，以预测哪些人可能会发表批评政府的言论，或是有所行动。

Such technology, if perfected, would give authoritarian governments a powerful tool to use against perceived enemies.

这种技术一旦完善，将为威权政府提供一个对付假想敌的强大工具。

The idea that an authoritarian government would use artificial intelligence to suppress dissent is troubling enough. But the use of A.I. to predict dissent well before a person has taken action has become a nightmare scenario, according to some involved in the industry.

威权政府利用人工智能镇压异见的想法已经足够令人不安。但一些业内人士表示，在个人采取行动之前就使用人工智能来预测异见倾向已经成为一种噩梦般的场景。

“This is what happens when mass surveillance meets A.I.,” said Brett J. Goldstein, the director of the Wicked Problems Lab at Vanderbilt’s Institute of National Security. “Without checks and balances, what China is doing to its own citizens is a preview of what becomes possible anywhere these tools go unchecked.”

“这就是大规模监控与人工智能结合时会发生的情况，”范德比尔特大学国家安全研究所棘手问题实验室主任布雷特·戈德斯坦说。“在没有制衡的情况下，中国对其本国公民所做的事情预示着，任何缺乏监管的地方都可能发生这样的事情。”

The Vanderbilt researchers found that Geedge, working with its government-supported research arm, MESA Lab, was developing technology that would generate profiles of Chinese citizens and then use A.I. to highlight who may pose a political risk.

范德比尔特大学的研究人员发现，积至正与其政府支持的研究分支MESA实验室合作，开发一种技术，能生成中国公民的个人档案，然后使用人工智能来筛选出可能构成政治风险的人员。

But the company’s progress appears to have been hampered by Biden-era export controls on U.S.-designed computer chips that power artificial intelligence. That suggests that U.S. restrictions may have slowed China’s development of the next generation of surveillance technology.

但拜登政府时期对驱动人工智能的美国设计计算机芯片实施的出口管制似乎阻碍了该公司的进展。这表明，美国的限制措施可能已经减缓了中国下一代监控技术的发展。

Documents from Geedge show that in 2024, when tough U.S. export controls on China were in place, the company and its lab struggled to find enough computing power for surveillance technology.

积至的文件显示，2024年，当美国对中国实施严厉的出口管制时，该公司及其实验室很难为监控技术找到足够的算力。

英伟达首席执行官黄仁勋，这是一家总部位于美国的科技公司，设计全球最强大的人工智能芯片。

Brett V. Benson, a political science professor at Vanderbilt, said the documents refer to GPU limitations. GPUs — graphic processing units — are the chips that power A.I. models. As a result of the limitations, he said, the company began using older, less powerful A.I. models and chips.

范德比尔特大学政治学教授布雷特·本森表示，这些文件提到了 GPU（图形处理器）的限制。GPU是驱动人工智能模型的芯片。他说，由于这些限制，该公司开始使用较旧、性能较低的人工智能模型和芯片。

Today, Geedge has access to enough graphic processing units for its current products, according to U.S. officials. But to carry out the most ambitious version of its predictive technology, the company would probably need more advanced chips than China can acquire, the officials said.

美国官员表示，如今，积至可以获得足够的图形处理器用于其当前的产品。但官员们指出，要实现其最具野心的预测技术版本，该公司可能需要比中国所能获得的更先进芯片。

Former Biden administration officials said intelligence showed that their policies restricting China from high-end, U.S.-designed chips had worked at keeping the U.S. artificial intelligence edge and at slowing the development of other Chinese technology.

前拜登政府官员表示，情报显示，他们限制中国获取美国设计的高端芯片的政策在保持美国人工智能优势和减缓中国其他技术发展方面发挥了作用。

But whether U.S. export controls will slow China’s development of more oppressive uses of artificial intelligence is an open question.

但美国的出口管制是否会减缓中国开发更具压迫性的人工智能应用仍是一个悬而未决的问题。

The Trump administration has weakened some of the Biden-era export controls and kept restrictions on the most advanced processors from Nvidia, the U.S. company that designs the world’s most powerful artificial intelligence chips. During President Trump’s recent trip to Beijing, U.S. officials said that China would have access to a more advanced version of the Nvidia chips.

特朗普政府放宽了拜登时代的一些出口管制，保留了对英伟达最先进处理器的限制。英伟达是设计全球最强大人工智能芯片的美国公司。在特朗普总统最近的北京之行中，美国官员表示，中国将可以获得更先进版本的英伟达芯片。

But China is trying to wean its artificial intelligence companies off U.S.-designed chips, so that export controls can no longer constrain their ambitions.

但中国正试图让其人工智能公司摆脱对美国设计芯片的依赖，以便出口管制无法再束缚它们的雄心。

在特朗普总统最近访问北京期间，美国官员表示，中国将能够获得英伟达芯片的更先进版本。

Profiling Future Critics

为未来的批评者建立档案

The new information was gleaned from a trove of data, including 100,000 documents from Geedge that were originally posted last September. Working with the documents, Wired and other publications have outlined how Geedge exported its network security software to countries including Ethiopia, Kazakhstan, Myanmar and Pakistan, enabling them to conduct mass surveillance on mobile networks.

这些新信息是从大量数据中收集的，其中包括最初在去年9月发布的10万份积至文件。通过研究这些文件，《连线》(Wired)和其他出版物概述了积至如何将其网络安全软件出口到埃塞俄比亚、哈萨克斯坦、缅甸和巴基斯坦等国家，使这些国家能够对移动网络进行大规模监控。

Vanderbilt researchers scouring the documents focused not on the technology Geedge was selling, but the next generation of technology it was trying to develop, as well as the constraints that could prevent it from putting the strategy into effect.

范德比尔特大学的研究人员在梳理这些文件时，关注的重点并非积至正在销售的技术，而是它试图开发的下一代技术，以及可能阻碍其实施战略的限制因素。

In the first months of 2024, according to the Geedge documents, the company’s researchers were working to develop behavioral profiles of people based on telecommunications, social media and location data. A.I. models were used to classify people and to “detect harmful information,” often a euphemism used by the Chinese Community Party to identify political dissent or other material the government wants suppressed.

根据积至的文件，在2024年的前几个月，该公司的研究人员致力于根据电信、社交媒体和位置数据开发人们的行为档案。人工智能模型被用来对人群进行分类，并“检测有害信息”——这是一种委婉说法，通常是指中国共产党用来识别政治异见或其他政府希望压制的材料。

In one meeting on Feb. 5, 2024, researchers discussed how to build profiles of people to “identify their intent” and “achieve discovery of harmful information,” according to minutes of the meeting.

根据2024年2月5日某次会议的纪要，研究人员讨论了如何建立人员档案以“识别其意图”并“实现对有害信息的发现”。

The Geedge researchers appeared to be developing tools to use artificial intelligence to predict who could become critics of the Chinese government, based on the data patterns the company’s surveillance technology collected.

积至的研究人员似乎正在开发一种工具，利用人工智能根据该公司监控技术收集的数据模式，预测哪些人可能会成为中国政府的批评者。

中科天玑致力于开发一款基于人工智能的软件，用于传播针对性的中国宣传内容。

“Geedge’s research team was doing more than just documenting behavioral patterns. They were trying to predict what citizens might do next and with whom,” Mr. Benson said. “Those stockpiles of data on ordinary materials are raw materials for generating profiles that determine who you are and what you will do next.”

“积至的研究团队所做的不只是记录行为模式。他们还试图预测公民接下来可能会做什么，以及和谁一起做，”本森说。“那些关于普通材料的庞大数据储备，是生成个人档案的原材料，这些档案可以决定你是什么样的人，以及你下一步会做什么。”

Geedge did not respond to a request for comment.

积至没有回应置评请求。

The documents suggest that the company’s team was working to link the physical movements with other online activity, including what movies people watched and what books they read, according to the Vanderbilt researchers.

范德比尔特大学的研究人员表示，这些文件表明，该公司的团队正致力于将物理移动轨迹与其他网络活动联系起来，包括人们看什么电影和读什么书。

Geedge’s research was in some ways similar to work by another Chinese company with ties to the state, GoLaxy. Last year, Vanderbilt and The New York Times outlined GoLaxy’s efforts to develop artificial intelligence-powered software that would push targeted propaganda — themes supported by the Chinese government and against views that Beijing opposed.

积至的研究在某些方面与另一家与政府有联系的中国公司中科天玑的工作相似。去年，范德比尔特大学和《纽约时报》概述了中科天玑开发人工智能驱动软件的努力，该软件将推送定向宣传，这些宣传主题得到中国政府的支持，并且反驳北京所反对的观点。

In some ways, Geedge’s work also evokes the 2002 movie “Minority Report,” in which the government predicts who may commit a crime, and a specialized police unit moves to arrest people before a violation occurs. While seen as infallible, the “precogs” who predict the future in the movie wrongly accuse the main character, played by the actor Tom Cruise, of a murder.

在某些方面，积至的工作也让人想起2002年的电影《少数派报告》(Minority Report)，在电影中，政府能预测谁可能会犯罪，一个专门的警察部队会在违法行为发生前采取行动逮捕这些人。尽管被视为绝无差错，但电影中预测未来的“先知”却错误地指控了由演员汤姆·克鲁斯扮演的主角涉嫌谋杀。

China has been pursuing predictive surveillance technology for years. But as the Geedge documents show, artificial intelligence models have sped the development of such tools.

中国多年来一直在追求预测性监控技术。但正如积至的文件所示，人工智能模型加速了此类工具的开发。

And Geedge shows how predictive policing technology could soon become a Chinese export.

而且积至展示了预测性警务技术可能很快成为中国出口产品的趋势。

DeepSeek的手机端人工智能助手。

Fine-Tuning a Surveillance State

完善监控国家体系

While the documents do not cover recent months, there is no evidence that Geedge has finalized or deployed the predictive technology, according to U.S. officials who spoke on condition of anonymity to discuss sensitive assessments.

尽管这些文件没有涵盖最近几个月的情况，但据要求匿名讨论敏感评估的美国官员称，没有证据表明积至已经定型或部署了这种预测技术。

But officials acknowledged that Chinese firms were working toward such technology to fine-tune their surveillance state. And China’s Public Security Bureaus have been racing to use DeepSeek, the country’s leading artificial intelligence model, to pursue other predictive policing technology, according to experts and government officials.

但官员们承认，中国企业正在努力开发这种技术，以完善其监控国家体系。据专家和政府官员透露，中国的公安局也一直在竞相使用该国领先的人工智能模型DeepSeek，开发其他预测性警务技术。

U.S. officials said that Geedge and other Chinese surveillance companies probably have enough computing power for their current-generation tools.

美国官员表示，积至和其他中国监控公司可能拥有足够的算力来支持其当前一代的工具。

But predictive tools, particularly if the company eventually incorporates intercepted phone calls or surveillance video, could come up against constraints on computing power.

但预测性工具——尤其是如果该公司最终将截获的电话通话或监控视频整合进来——可能会面临算力上的限制。

If the company tried to tap into those kinds of surveillance streams to predict security threats, it would more quickly come up against the limits of China’s current computing power, American officials said.

美国官员表示，如果该公司试图利用这类监控数据流来预测安全威胁，将更快地触及中国当前算力的极限。

“Chinese security services are dealing with an overload of data,” said Jimmy Goodrich, a senior fellow at the University of California’s Institute on Global Conflict and Cooperation. “They have too much information at once. The real value of artificial intelligence is that they can triage the data and find the threats. But their ability to scale that depends on their access to compute. This is what makes export controls so important.”

“中国安全部门正面临数据过载的问题，”加州大学全球冲突与合作研究所高级研究员吉米·古德里奇说。“他们同时拥有太多的信息。人工智能的真正价值在于他们可以对数据进行分类并发现威胁。但他们扩大这种能力取决于他们获得算力的途径。这就是出口管制如此重要的原因。”