OpenAI Dismantles Russian 'Bad Grammar' Op That Used AI to Spam Telegram

OpenAI has banned a cluster of accounts tied to a Russian influence operation it calls “Bad Grammar,” a network that built a crude AI-powered pipeline to spam political comments on Telegram. The operation’s name is a jab at its own sloppy output—operators repeatedly generated comments riddled with grammatical errors while targeting audiences in Russia, Ukraine, the U.S., Moldova, and the Baltic states. This wasn’t a deepfake crisis. It was something much more mundane and, in a way, revealing about the current state of AI-enabled information warfare. The group primarily used OpenAI’s models to debug code for automated posting and to churn out short replies to specific Telegram channels, with the pro-Russia @Slavyangrad being its most frequent target.

What’s striking is the network’s sheer lack of traction. Despite deploying at least a dozen Telegram accounts and generating comments in the voice of fake American personas from both ends of the political spectrum, the content was mostly a dud. OpenAI’s assessment puts the operation at Category 1 on the Breakout Scale—the lowest rung, defined by activity on a single platform with no meaningful amplification from real people. The comments, which argued against U.S. support for Ukraine or accused Ukrainian and Moldovan leaders of corruption, typically received zero likes or replies. The most bizarre detail? Some generated texts suggest the operators accidentally got into a conversation with a cryptocurrency scammer. As an industry reporter, I find this more tragicomic than terrifying.

The tactics, however, are a familiar playbook given an automated twist. The “two-faced” approach—posing as conflicting voices to sow division—has been seen in Russian operations for years. After the Moscow terrorist attack in March, the network pivoted quickly, spinning up new accounts focused on Tajikistan and pushing messages of unity with Russia. It’s a classic reactive propaganda move, but automating it with generative AI didn’t make it any more effective. The real story here might be that AI-generated spam, when dumped into a platform like Telegram without a sophisticated distribution strategy, is still just spam. It’s the digital equivalent of shouting into a void.

This takedown underscores a gap between the fear of AI-powered disinformation and the messy reality. While the technique of generating bulk commentary is novel, the impact assessment shows the network struggles to build an audience. The operation failed to drown out other views or drive any conversation. It’s a useful data point for a debate often dominated by hypothetical doomsday scenarios. The Bad Grammar operation didn’t fail because the AI was bad; it failed because the operation itself, much like its ungrammatical English, simply wasn’t good enough to matter.