OpenAI Disrupts 'Data Center Bandwagon' Influence Operation
OpenAI has banned a cluster of ChatGPT accounts linked to a private Chinese technology company that conducted covert influence operations targeting the United States. The actors used ChatGPT to generate deceptive social media content, create fake personas, and develop operational strategies to evade detection on platforms like X and Facebook.
China-Linked Actor and Operational Methods
The influence operation was conducted by a cluster of accounts likely originating in China, operating via VPNs to bypass regional access restrictions. The actors are believed to be part of a social media operations team at a private Chinese technology company working for provincial-level government clients, aligning with a commercial ecosystem that supports Chinese Party-state priorities in public opinion guidance.
Key operational details include:
- Persona Fabrication: Actors posed as Americans from various backgrounds, including students, mothers, clerks, and investors, to appear as organic US-based users.
- Language Usage: Prompts were submitted in Simplified Chinese, with requests for outputs in both English and Chinese.
- Workflow Automation: ChatGPT was used to write code for automating logins and managing interactions across multiple social media platforms, as well as for text processing tasks like extracting usernames and formatting data for worksheets.
Targeted Narratives and Influence Tactics
The campaign targeted two primary audiences with distinct narratives designed to manipulate public opinion:
US-Based Audiences
Actors generated English-language comments and images claiming that AI applications and data centers were driving up electricity demand and increasing costs for ordinary American households. This included:
- Contextualized Misinformation: Using legitimate regional news reports about power grid capacity auctions to frame AI-driven power demand as a cause of rising electricity prices.
- Visual Manipulation: Using ChatGPT to edit generic electricity market images by adding text to support the narrative that ordinary people were subsidizing AI infrastructure.
- Social Media Distribution: Posting content on X using hashtags such as #capacityauction, #datacentersuccess and #datacenters.
Overseas Chinese and Dissidents
The operation sought to harass Chinese dissidents and political commentators, including Li Ying (Teacher Li), Lu Yiheng, Xu Chi, and the X account @SydneyDaddy1. While the actors attempted to use ChatGPT to generate insulting comments directed at Li Ying's team, OpenAI's models refused to generate these inflammatory personal attacks.
Platform Evasion and Strategic Planning
The actors used ChatGPT to polish work reports that detailed their strategies for maintaining long-term account viability and evading platform detection systems. Their operational security focused on:
- Persona Development: On Facebook, the actors emphasized building "real, trustworthy and daily life personas" using lifestyle and professional content to create a brand before amplifying narratives.
- Organic Engagement Simulation: Using cross-account interactions to simulate organic engagement to preserve the appearance of authenticity.
- Facebook Ecosystem Analysis: The group analyzed Facebook's groups, pages, hashtags, advertising tools, and recommendation systems to implement a "dual-track approach" combining organic engagement with paid Facebook ads.
- Risk Mitigation: Creating backup accounts and separating operational activity to avoid detection of coordinated behavior by platform security systems.