OpenAI has dismantled a sophisticated influence campaign it dubbed “Trolling Stone,” where bad actors used ChatGPT to flood social media with pro-Russian propaganda. The operation centered on defending Konstantin Rudnev, an alleged cult leader arrested in Argentina, by generating thousands of comments and fake news articles. While the content targeted Argentine audiences, the digital fingerprints pointed elsewhere. Investigators traced the activity to a disjointed network of users in Pakistan, Armenia, Uruguay, Argentina, and Kazakhstan. This geographic scatter suggests a for-hire model rather than a single centralized command center, with the Pakistan-based node likely acting as a commercial contractor.
The scheme relied on classic astroturfing tactics supercharged by generative AI. Operators created Facebook Pages posing as legitimate Argentine news outlets, all registered in mid-October 2025. These pages published short, AI-generated news briefs and translated Russian propaganda into local Spanish. To avoid detection, some users manually stripped em-dashes from the text, a subtle tell that they were trying to mask the machine-generated origin of the prose. The volume was high; internal instructions demanded each fake account post 20 times daily, with strict requirements for progress reports. This wasn’t just random spam. It was a coordinated effort to manufacture consensus.
What made this campaign particularly brazen was its direct harassment of established media. When Rolling Stone Argentina published a critical piece calling Rudnev an “extravagant guru,” the network pivoted. Accounts generated comments accusing the magazine of journalistic failure, posting them on unrelated music threads to maximize visibility. In one instance, operators even used ChatGPT to edit a supposed “refutation” article by a fake Spanish-speaking journalist before publication. OpenAI noted that some prompts explicitly asked the model to generate content that would slip past Meta’s moderation guidelines. This reveals a growing sophistication in adversarial prompting, where users treat safety filters as puzzles to be solved rather than rules to be followed.
The dispersion of this network highlights a shifting landscape in information warfare. By outsourcing parts of the operation to different countries, the architects likely hoped to dilute attribution and complicate enforcement. Yet the linguistic slips and behavioral patterns gave them away. As AI tools become more accessible, we can expect more hybrid operations that blend human coordination with automated content generation. The real challenge isn’t just detecting the AI text, but identifying the human hands guiding the prompt engineering. OpenAI’s intervention stopped this specific burst, but the playbook remains available for anyone willing to pay the price.
