OpenAI Operation Peer Review: Disrupting AI-Assisted Surveillance Planning
OpenAI has banned a cluster of ChatGPT accounts, dubbed "Operation Peer Review," which were likely originating in China and used to assist in surveillance planning, political research, and the development of social media monitoring tools. This action disrupts a network that utilized AI for capability development and internal analysis rather than large-scale public influence operations.
Network Identification and Behavioral Patterns
The "Peer Review" network consisted of ChatGPT accounts that exhibited behavioral patterns consistent with mainland Chinese business hours and used Chinese-language prompting. OpenAI noted that the volume and variety of the activity were consistent with manual prompting rather than automated API usage. In at least one instance, evidence suggested a single account was shared by multiple operators.
Primary Operational Workstreams
The operators within this cluster engaged in four primary categories of activity:
Political and Government Research
The network used ChatGPT as a basic research tool to gather publicly available information on United States think tanks, as well as politicians and government officials in the United States, Australia, and Cambodia.
Document Analysis and Translation
Operators used the models to read, translate, and analyze screenshots of English-language documents. These documents included announcements for Uyghur rights protests in various Western cities—potentially sourced from social media—and materials concerning diplomatic and government topics in the Indo-Pacific region. OpenAI stated there was insufficient evidence to determine the authenticity of these documents or how they were obtained.
Content Generation for Influence
The actors generated short- to medium-length comments regarding US policies, politics, and Chinese dissident organizations, specifically the Falun Gong. Some prompts requested the model to assume the persona of an English-speaker named "Thompson." However, OpenAI could not identify any of these generated comments being posted online.
Surveillance Tool Development
The network used ChatGPT to edit and debug code and generate promotional materials for an AI-powered social media listening tool called the "Qianyue Overseas Public Opinion AI Assistant" (千阅境外舆情AI助手).
According to the tool's descriptions, which OpenAI could not independently verify, the assistant was designed to:
- Ingest and analyze posts and comments from platforms including X, Facebook, YouTube, Instagram, Telegram, and Reddit.
- Identify social media conversations regarding Chinese political and social topics, specifically calls to attend human rights demonstrations in China.
- Feed resulting insights to Chinese authorities, including intelligence agents monitoring protests in the US, Germany, and the UK, as well as Chinese embassies abroad.
Technical Implementation and Model Usage
While the network used ChatGPT for development and refinement, the actual execution of the surveillance tool did not appear to be powered by OpenAI models. Code debugged by the network frequently referenced Meta's Llama 3.1 8B deployed via Ollama as the analysis and generation driver.
Additionally, the network debugged code intended for malware analysis and referenced other models, including Alibaba Cloud's Qwen and an unspecified model from DeepSeek. One operator also claimed in a performance review that they had generated phishing emails for clients in China, though OpenAI found no evidence that this email activity was powered by AI.
Assessment of Impact
OpenAI reports that the majority of the activity was focused on internal capability development—such as code debugging and image analysis—rather than public distribution. While some content resembled social media comments, OpenAI was not able to find evidence of these comments being posted online. OpenAI noted that assessing the full impact of this operation would require collaboration with other stakeholders, including operators of the open-source models used by the network.