OpenAI Disrupts "Bad Grammar" Russian-Linked Influence Operation

OpenAI has banned a network of accounts linked to individuals in Russia that used AI models to automate the generation of political comments for distribution on Telegram. This operation, dubbed "Bad Grammar" due to its repeated use of ungrammatical English and difficulty in building an audience, represents an attempt to use generative AI to scale influence operations (IO).

Technical Pipeline for Comment Spamming

The "Bad Grammar" network established a comment-spamming pipeline that integrated AI model usage with automation tools. The process followed three primary stages:

  1. Code Debugging: Operators used OpenAI models to debug code designed to automate the posting process on Telegram.
  2. Content Generation: The models were used to generate comments in both Russian and English, specifically tailored to reply to target posts.
  3. Distribution: The network utilized at least a dozen Telegram accounts to post the generated content.

Target Audiences and Persona Development

The network targeted audiences in Russia, Ukraine, the United States, Moldova, and the Baltic States. The approach to content generation differed by language:

  • English Content: Operators created fake personas representing various demographics and political spectrums within the United States to simulate diverse domestic viewpoints.
  • Russian Content: Instructions were more thematic and less focused on specific personas.

Activity was heavily concentrated on a small number of Telegram channels. The most-mentioned channel was the pro-Russia channel @Slavyangrad, followed by @police_frequency and @SGTNewsNetwork (a channel Meta previously reported as originating in Iran).

Political Themes and Content Strategy

The campaign focused on several key political themes across different languages:

  • War in Ukraine: A primary topic in both languages.
  • Regional Stability: The political situations in Moldova and the Baltic States.
  • United States Politics: English-language comments focused on immigration, economic hardship, and current events to argue against U.S. support for Ukraine.
  • Russian-language comments: These focused on accusing the presidents of Ukraine and Moldova of corruption and betraying their people to Western "interference."

Following the March 22 terrorist attack in Moscow, the network expanded its activity to include a new cluster of Telegram accounts focused on Tajikistan, calling for unity between Tajiks and Russians and respect for Russian law.

Influence Operation Impact Assessment

Despite the use of generative AI, the "Bad Grammar" network had minimal real-world impact. OpenAI assesses the operation as Category 1 on the Breakout Scale—the lowest level of impact—due to the following factors:

  • Low Engagement: Very few comments received likes or replies.
  • Limited Reach: The network's comments typically represented a minority of the replies to any one post, failing to drown out other perspectives.
  • Limited Amplification: There was no evidence of significant amplification by users outside the network.

Additionally, OpenAI noted that the network occasionally used models to generate private messages, which appeared to be interactions with a cryptocurrency scammer.

Sources