OpenAI Disrupts Operation Sneer Review China-Origin Influence Activity

OpenAI has disrupted a covert influence operation of Chinese origin, known as Operation Sneer Review, by banning ChatGPT accounts used to generate bulk social media content and internal organizational documents. This activity represents an attempt to use generative AI to simulate organic engagement and spread narratives aligned with China's geostrategic interests across multiple global platforms.

Actor and Operational Intent

OpenAI identified and banned accounts that were using ChatGPT to produce social media posts and internal performance reviews. The actor's intent was to conduct a covert influence operation (IO) by generating content in English, Chinese, and Urdu.

Key characteristics of the actor include:

  • Language and Origin: Prompts were primarily issued in Chinese and focused on geopolitical topics relevant to China. One user explicitly claimed in a prompt to work for the Chinese Propaganda Department, though OpenAI noted it has no independent evidence to verify this specific claim.
  • Operational Use Cases: Beyond public-facing content, the actors used ChatGPT to draft internal policy documents, an essay written in the style of an official public security document, and performance reviews detailing the steps taken to manage the operation.

Behavioral Patterns and Platform Distribution

Operation Sneer Review employed a specific pattern of "main" accounts and supporting accounts to create a false impression of organic engagement.

Social Media Strategy

  • Engagement Simulation: A typical pattern involved a main account posting initial content, followed by a series of reply comments generated by other network accounts to simulate a conversation.
  • TikTok: The network used screen names in various languages and alphabets (e.g., Korean, Thai, Hebrew) that were often unrelated to the language of the content being posted (e.g., an account with a Korean name posting in Urdu).
  • X (formerly Twitter): Accounts typically utilized cartoon profile images and some used names associated with cryptocurrency.
  • Facebook: Activity was limited to creating two Pages posing as news outlets; however, these pages had no followers or likes.
  • Reddit and Web Forums: The network generated longer-form posts, including one on Reddit and two on other web forums.

Targeted Content and Narratives

The content generated by the network was strictly aligned with China's geostrategic interests, focusing on three primary themes:

Taiwan and "Reversed Front"

The operation's namesake, "Sneer Review," comes from its activity targeting the Taiwan-centric game Reversed Front (逆統戰), which imagines resistance against the Chinese Communist Party. The network generated dozens of critical comments and a long-form article claiming the game faced widespread backlash.

Pakistani Activism

Content in English and Urdu targeted Pakistani activist Mahrang Baloch, who has criticized Chinese investments in Balochistan. The network used a TikTok account and Facebook Page to post a video falsely accusing Baloch of appearing in a pornographic film, supported by hundreds of generated comments to simulate engagement.

U.S. Foreign Aid

The network generated content on TikTok and X regarding the closure of the U.S. Agency for International Development (USAID), producing comments that either praised or criticized the move.

Impact Assessment

OpenAI assesses that Operation Sneer Review was in its infancy when it was disrupted. The actual reach of the operation was limited, with Facebook Pages having zero followers and Reddit posts being largely blocked or removed.

Engagement Metrics

  • TikTok: Two videos combined for 25,000 likes.
  • X: Main account tweets typically received approximately 10,000 views each.

OpenAI cautions that these figures should be treated with skepticism because many of the comments were generated by the network itself, indicating significant inauthentic engagement. Using the Brookings Institution's Breakout Scale for influence operations, OpenAI assesses the activity as being at the low end of Category 3, assuming engagement figures were authentic, but notes this would be revised downward if more inauthentic engagement is proven.

Sources