Bad Grammar Gave It Away: OpenAI Disrupts a Russia-Linked AI Influence Operation

Hi, it's Shiichan! Today I found a report from OpenAI about an attempt to misuse AI that got shut down, so let me walk you through it.

OpenAI News openai.com

What was announced?

OpenAI News announced that it disabled a network of accounts tied to a previously unreported, Russia-linked covert influence operation. The team nicknamed it "Bad Grammar" because the posts were, well, full of grammar mistakes — which made it hard for the operation to get any real engagement in the first place.

Why it matters

The Bad Grammar operation is a real-world example of generative AI being used as a tool for state-linked influence operations. The operators first used OpenAI's models to debug code designed to automate posting on Telegram, then used the models again to generate comments in both English and Russian tailored to ongoing Telegram discussions. From writing the automation code to writing the actual comments, AI was used at every step.

What changes

OpenAI banned the accounts behind this operation, cutting off the automated posting and AI-generated comments distributed through those Telegram accounts. Publishing the specific tactics also gives other platforms and researchers a clearer picture of what AI-generated, multilingual comment spam looks like, making it easier to spot similar campaigns going forward.

Dive Deep

• At least a dozen Telegram accounts were used to spread the generated comments
• Main targets included the pro-Russia channel @Slavyangrad, plus English-language channels @police_frequency and @SGTNewsNetwork
• Topics ranged from the war in Ukraine, Moldovan politics, and developments in the Baltic states, to U.S. domestic issues like immigration and the economy
• Russian-language comments attacked Ukrainian and Moldovan leadership, while English-language comments pushed an isolationist "stop supporting Ukraine" message — the two languages carried different messaging
• The operation even ran a "two-faced" trick, creating opposing personas to manufacture the appearance of a genuine debate
• OpenAI rated the operation as Category 1 on the Breakout Scale, its lowest tier for measuring how far an influence operation spreads. In practice, most of the generated comments got little to no engagement and represented only a minority response to the posts they targeted

Wrap-up

• OpenAI disabled a Russia-linked influence operation network nicknamed "Bad Grammar"
• The operation used AI end-to-end, from debugging Telegram automation code to generating comments in English and Russian
• Targets included at least a dozen Telegram accounts such as @Slavyangrad, covering Ukraine, Moldova, the Baltic states, and U.S. domestic politics
• Its reach was rated Category 1 (the lowest tier) on the Breakout Scale, and real engagement was minimal

This report is a great look at both how AI gets misused for influence operations and how ineffective that misuse can actually be — worth a read if you care about AI safety and how information operations really work.