PRC-linked abuse: Surveillance and influence activity

This case study was originally published in OpenAI’s October 2025⁠(opens in a new window) report.

Individuals apparently linked to PRC government entities using ChatGPT to assist in development, profiling and bureaucratic functions.

As we laid out in our submission to the Office of Science and Technology Policy’s U.S. AI Action Plan in March, we believe that making sure AI benefits the most people possible means enabling AI through common-sense rules aimed at protecting people from actual harms, and building democratic AI. By democratic AI, we mean AI that is shaped by the democratic principles America has always stood for. This includes preventing the use of AI tools by authoritarian regimes to amass power and control their citizens.

Our investments in detecting and disrupting attempts by authoritarian regimes to abuse ChatGPT’s capabilities not only protect people now, but provide important insights into how those regimes might abuse future capabilities. This section discusses one set of such findings concerning individuals potentially linked to such regimes, focused on the PRC.

As we wrote in June, the PRC is making real progress in advancing its autocratic version of AI. Our disruption of ChatGPT accounts used by individuals apparently linked to Chinese government entities shines some light on the current state of AI usage in this authoritarian setting. Some elements of this usage appeared aimed at supporting large-scale monitoring of online or offline traffic, underscoring the importance of our ongoing attention to potential authoritarian abuses in this space. Other elements of this usage appeared aimed at more in-depth profiling. We banned the accounts associated with this activity and are sharing insights in this report.

This study is based on a selection of our disruptions of violating activity that we identified and banned. That activity was consistent with individual users using ChatGPT, rather than large-scale, institutional adoption of our models. As such, it is a limited snapshot of the usage of different AI models in this context.

Some of the accounts that we banned appeared to be attempting to use ChatGPT to develop tools for large-scale monitoring: analyzing datasets, often gathered from Western or Chinese social media platforms. We reported in February on one such case, aimed at designing an AI-powered social media listening tool the operators claimed was for the Chinese security forces. These users typically asked ChatGPT to help design such tools or generate promotional materials about them, but not to implement the monitoring.

For example, we recently banned a user, possibly using a VPN to access our services from China, who was asking ChatGPT to help design promotional materials and project plans for a social media listening tool, purportedly for use by a government client. The tool was described as a social media ‘probe’ (探针) that could allegedly scan Twitter/X, Facebook, Instagram, Reddit, TikTok, and YouTube for what the user described as extremist speech, and ethnic, religious, and political content. The user did not use our model to conduct the actual social media monitoring. We are unable to independently verify if this tool has been used by a Chinese government entity.

Similarly, we banned a second user, likely connected to a government entity, who asked ChatGPT to help write a proposal for what they described as a High-Risk Uyghur-Related Inflow Warning Model (高危涉维吾尔关注人员流入预警模型). The proposal described a tool that would analyze transport bookings and compare them with police records in order to provide early warning of travel movements by people classed as Uyghur-related and high-risk, an otherwise undefined category. The user did not ask our model to help build such a tool, only to develop a general proposal; as such, we are unable to independently confirm what technology they intended to use or whether such a tool has been used by a Chinese government entity.

We also banned users who appeared to be linked to Chinese government entities and who appeared to be trying to use ChatGPT for more bespoke, targeted profiling and online research.

For example, one user asked ChatGPT to identify funding sources for an X account that criticized the Chinese government. A second user asked ChatGPT to identify the organizers of a petition in Mongolia. In both these cases, our models only returned publicly available information, which did not include sensitive details such as funding sources or the identities of petition organizers.

Some of this activity used our models as an open-source research tool, in a similar way to which earlier users might have used internet or social media searches. For example, a third user asked ChatGPT to identify and summarize daily breaking news that would be of relevance to China, including on sensitive topics such as the anniversary of the Tiananmen Square massacre in 1989 and the birthday of the Dalai Lama.