Chinese spy services have invested heavily in artificial intelligence to create new tools to speed analysis, provide early warning of threats and potentially help shape operational plans during a war, according to a new report.

据一份新的报告称，中国的情报机构已在人工智能领域投入巨资，创建新的工具来加快分析速度，提供威胁预警，并有可能在战争期间帮助制定作战计划。

China, like the United States, hopes that artificial intelligence will improve the efficiency and accuracy of its intelligence analysis, allowing it to collect more intelligence and analyze it faster and more cheaply.

与美国一样，中国也希望人工智能能够提高情报分析的效率和准确性，使其能够收集更多情报，并且更快、更便宜地进行分析。

The study, by Recorded Future’s Insikt Group, which studies cybersecurity and other threats from nation-states, terrorists and criminal groups, comes amid rising concern about how Chinese spy agencies will use A.I. to power covert actions, as Western intelligence services also embrace the technology.

这份研究报告由“记录未来”的Insikt小组撰写，该组织研究网络安全以及来自民族国家、恐怖分子和犯罪团伙的其他威胁。这项研究是在外界对中国间谍机构将如何利用人工智能技术增强秘密行动能力的担忧日益加剧之际进行的，与此同时，西方情报机构也在采纳这项技术。

The researchers reviewed patent applications by the People’s Liberation Army, publicly available contracts and other material to better understand how China’s military and intelligence services have invested in artificial intelligence.

研究人员对中国人民解放军的专利申请、公开合同和其他材料进行了评估，以便更好地了解中国军方和情报部门如何投资于人工智能。

Recorded Future found that China is probably using a mix of large language models, technology that can analyze huge amounts of data and communicate its results in human language. Meta and OpenAI are thought to be among the American models that China is using, along with Chinese models from DeepSeek, Zhipu AI and others.

“记录未来”发现，中国可能正在混合使用各种大语言模型，这种技术可以分析大量数据，并用人类语言传达结果。Meta和OpenAI被认为是中国正在使用的美国模型之一，此外还有DeepSeek、智谱AI等公司的中国模型。

The C.I.A. and other American spy agencies have stepped up their use of artificial intelligence, both to improve analytic work and to help overseas operatives remain undiscovered. One tool developed by the C.I.A. is designed to help analysts assess the positions of foreign leaders, creating virtual versions of the officials that are powered by artificial intelligence.

中央情报局和其他美国情治单位已经加强了对人工智能的使用，一方面是为了改进分析工作，另一方面是为了帮助海外特工不被发现。CIA开发的一种工具旨在帮助分析人员评估外国领导人的立场，通过人工智能创建这些官员的虚拟版本。

The Pentagon announced on Monday that it was awarding a $200 million contract to OpenAI. The company, in a release, said its OpenAI for Government initiative would be used to improve administrative operations, including health care, but also improve work on military acquisition programs and support “proactive cyberdefense.”

周一，五角大楼宣布，将授予OpenAI一份价值2亿美元的合同。OpenAI在新闻稿中表示，其“政府OpenAI”计划将用于改善包括医疗保健在内的行政运营，同时也将改善军事采购项目的工作，并支持“主动网络防御”。

Former American intelligence officials have said China’s large population has long given it a potential advantage over U.S. spy agencies, but artificial intelligence could even the playing field. Generative A.I. models can scan huge amounts of collected communications intelligence and queue the most interesting information for human analysts to examine.

前美国情报官员表示，长期以来，中国庞大的人口使其在与美国情报单位的竞争中具有潜在优势，但人工智能可以平衡竞争环境。生成式人工智能模型可以扫描收集到的大量通信情报，并将最引人关注的信息排列出来，供人类分析师审查。

Some U.S. officials said China’s investment in artificial intelligence was of little surprise, given its potential to improve analytic assessments. But the Recorded Future report found specific examples of how China could be using large language models and generative A.I. to not just improve its intelligence analysis, but also help military commanders improve targeting and operational plans.

一些美国官员表示，考虑到人工智能在改进分析评估方面的潜力，中国在人工智能方面的投资并不令人意外。但“记录未来”的报告发现了一些具体的例子，说明中国如何利用大语言模型和生成式人工智能，不仅可以改善其情报分析，还可以帮助军事指挥官改进目标定位和行动计划。

In October, Ordnance Science and Research Academy of China filed a patent application to use various forms of intelligence to train a military model. The application talks about the ways the model could be used, such as by crafting operational plans and helping battlefield intelligence analysts analyze friendly and enemy forces, according to Zoe Haver, the author of the study and a senior threat intelligence analyst at the Insikt Group.

去年10月，中国兵器科学研究院提交了一份利用各种形式的情报来训练军事模型的专利申请。该研究报告的作者、Inskt小组的高级威胁情报分析师佐伊·哈弗表示，该申请讨论了大语言模型的使用方式，比如制定行动计划，帮助战场情报分析师分析友军和敌军。

“This was very broad ranging, and intended to be applicable across the intelligence cycle,” Ms. Haver said.

“涉及范围非常广泛，旨在适用于整个情报周期，”哈弗说。

Over the last two years, China has tightened control over information about what its military and intelligence agencies have obtained. So while Recorded Future was often able to see the military’s procurement of generative A.I. models and servers, it was not always clear how the technology would be used. But some Chinese contractors appeared to have grand ambitions.

在过去两年中，中国加强了对军队和情报机构所获信息披露的控制。因此，虽然“记录未来”时常能够看到军方采购生成式人工智能模型和服务器，但并不总是清楚这项技术将如何使用。不过，一些中国承包商似乎有着宏伟的抱负。

China’s military and intelligence agencies appear to have quickly pivoted from open source and Western A.I. to DeepSeek, which unveiled a model rivaling OpenAI’s model, ChatGPT, the day after Christmas. Global interest in DeepSeek’s model exploded in January. By the end of February, military procurement records appeared showing Chinese companies quickly taking up DeepSeek’s technology.

中国的军队和情报机构似乎已迅速从开源和西方人工智能转向DeepSeek，该公司在圣诞节后的第二天发布了一款与OpenAI的模型ChatGPT相抗衡的模型。今年1月，全球对DeepSeek模型的兴趣激增。到2月底，军事采购记录显示，中国企业正在迅速采用DeepSeek的技术。

At the same time, American firms have cracked down on China’s use of their models.

与此同时，美国公司也在打击中国使用其模型的行为。

This month, OpenAI reported that it had disrupted several operations most likely originating in China that had tried to use its artificial intelligence tools in malicious ways. The operations were a combination of influence campaigns and surveillance, according to OpenAI. One of them tried to use ChatGPT to generate comments on social media sites about the dismantling of the U.S. Agency for International Development.

本月，OpenAI报告称，它中断了数起很可能源自中国的行动，这些行动试图以恶意方式使用其人工智能工具。据OpenAI称，这些行动是影响力宣传活动和监视的结合。其中一项行动试图利用ChatGPT生成关于美国国际开发署解散的评论，供其在社交媒体网站上发表。

To train a model to provide meaningful insight on intelligence, a government needs to give it access to its intelligence data, which can be difficult while still keeping classified material secure.

要训练一个模型提供有意义的情报见解，政府需要让它访问自己的情报数据，同时又要确保机密材料的安全，这可能很困难。

Chinese intelligence products are often infused with the ideology of the ruling Communist Party. Ms. Haver said a model trained on such reports would produce intelligence biased in the same way. But whether the Chinese government sees that as a problem is another question.

中国的情报产品经常融入执政共产党的意识形态。哈弗说，接受过这类报告训练的模型也会产生同样带有偏见的情报。但中国政府是否认为这是个问题，那又是另一回事了。

“Some Chinese public security researchers are talking about ChatGPT being used for intelligence,” Ms. Haver said. “And they are worried about how objectivity, neutrality, neoliberalism and capitalistic values could infiltrate Chinese intelligence work if they use foreign models.”

“一些中国公共安全研究人员正在讨论ChatGPT被用于情报，”哈弗说。“他们担心，如果使用外国模型，客观、中立、新自由主义和资本主义价值观可能会渗透到中国的情报工作中。”