OpenAI Discloses China-Linked Actors Using ChatGPT to Orchestrate Anti-US AI Data Center Public Opinion, Evidential Sufficiency and Motives Spark Controversy

OpenAI's threat intelligence team disclosed two independent cyber activity clusters on June 11, 2025. They are believed to be linked to China and used ChatGPT to generate negative content targeting the construction of AI data centers in the United States. The goal was to amplify public opposition to rising electricity prices and tariff policies.

How Activities Used AI Tools to Generate Content

The first activity cluster was named "Data Center Bandwagoners." Operators input prompts asking ChatGPT to create images and social media comments, claiming that data center expansion drove up electricity bills for US residents. Another cluster focused on tariff issues, generating content stating that tariffs were a covert means for China and the US to vie for global technological dominance. The prompts specifically required that only former US President Donald Trump appear in the content, excluding Chinese President Xi Jinping.

These prompts were all written in Simplified Chinese and required output in both English and Chinese versions. Operators bypassed China's internet restrictions using VPNs, and accounts posted content on platforms like X and YouTube, posing as American users. OpenAI's tracing found that the first cluster originated from an unnamed Chinese tech company that had signed contracts with multiple local Chinese governments.

Technical Limitations and Traces of Content Generation

While AI models can rapidly generate images and text in bulk, the output often exhibited obvious flaws. Some images had rigid compositions, and the text carried overtly propagandistic tones, lacking the natural expression habits of native English speakers. These characteristics became key clues for OpenAI in determining the source of the activities.

Meanwhile, operators did not rely entirely on a single model. They mixed manual editing with multiple rounds of prompt iteration to reduce the risk of detection by platforms. However, this mixed approach still left traceable language patterns and timestamp regularities.

Actual Assessment of Dissemination Effectiveness

OpenAI used the Bookmarks breakout scale to score the two activities, giving them 1 and 2 points respectively. This score indicates that the activities appeared on only one or multiple platforms, without observed meaningful interaction from target audiences. Researchers noted that such influence operations often tend to attach themselves to existing domestic controversies to amplify their visibility, but this case did not achieve that effect.

Report author Ben Nimmo pointed out that this is a typical case of foreign influence operations exploiting domestic issues for manipulation. The actors attempted to shape negative public perception of AI infrastructure by impersonating American users.

Public Divergence on National Security and Evidential Sufficiency

After the incident was exposed, discussions quickly diverged. One side argued that such AI-powered opinion interventions pose a direct threat to US critical infrastructure construction, calling for stricter foreign influence assessments in supply chain reviews and data center site selection. The other side noted that existing evidence mainly comes from platform behavior patterns and linguistic features, lacking documents or directives directly attributed to the Chinese government, making it difficult to rule out the possibility of amplification driven by geopolitical narratives.

Both sides did not deny the fact that AI tools had been used for large-scale content production, but they clearly disagreed on whether "the influence had translated into substantial public pressure." OpenAI itself is planning to raise hundreds of millions of dollars to build data centers in the US, leading some observers to question the stance of its report.