OpenAI Disrupts 'Corrupt Comment' Influence Operation

OpenAI has banned a cluster of API activity associated with an influence operation dubbed "Corrupt Comment." This operation used OpenAI models to generate English-language comments on X (formerly Twitter) to criticize the Anti-Corruption Foundation (FBK), founded by the late Alexei Navalny, and its leadership.

Operation Mechanics and Behavior

The "Corrupt Comment" operation utilized OpenAI's API to generate English-language text that was subsequently posted by fake accounts on X. These accounts primarily targeted senior figures at the Anti-Corruption Foundation and Navalny's associates.

Key characteristics of the operation's behavior include:

  • Lack of Contextual Awareness: While the generated comments were posted as replies to existing posts—often posts written in Russian—there is no evidence that the models were used to read, analyze, or generate relevant responses to the content of those posts. The comments were pre-generated and not reactive to the social media content.
  • Account Profiles: The X accounts involved were mostly created in December 2023. These accounts typically had zero followers and received zero replies to their posts.
  • Visual Deception: Profile pictures used by these accounts were either scenery images copied from the internet or fake images created using Generative Adversarial Networks (GANs), a technique dating back to 2019.

Content Strategy: "Theme and Variations"

The generated content focused on criticizing the Anti-Corruption Foundation, specifically questioning its transparency. The operation employed a "theme and variations" strategy, where the same essential message was repeated across many different posts using slight variations in wording to avoid detection or create an illusion of diverse opinion.

Impact Assessment

OpenAI assesses the impact of this operation as limited. Most identified accounts had no followers and received no engagement (replies). Furthermore, the English-language replies generated by this network were typically outnumbered by Russian-language replies from third parties unrelated to the operation, suggesting that the network did not successfully drown out the conversation on the platform.

Sources