"Bad Grammar": Russian-linked Telegram comment activity
OpenAI ● Covered by 3 sources
OpenAI banned a Russian op, "Bad Grammar," using AI to spam Telegram comments. It targeted Ukraine, Moldova, Baltic states, and US politics automatically.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI's latest threat report introduces a name that sounds almost comic until you look at what it actually did: Bad Grammar. This is a previously undocumented, Russia-linked operation that used OpenAI's own models to churn out political comments on Telegram, in both English and Russian, aimed at stirring debate around Ukraine, Moldova, the Baltic states, and, notably, domestic US politics.
The mechanics were not especially glamorous. Rather than building slick disinformation sites or deepfake video, the operators leaned on AI to generate short-form comments and replies that could be dropped into ongoing Telegram threads, apparently using scripts or bots to automate the posting itself. The goal seems to have been volume and persistence over polish: get enough AI-written opinions into enough feeds that the conversation tilts, even slightly, in a preferred direction.
What stands out is the geographic spread. Pairing Ukraine and Moldova with the Baltics and the United States suggests an operation less interested in a single narrative and more in probing multiple fronts where Russian state interests already run hot, from NATO's eastern flank to America's polarized election-year discourse. That breadth, done cheaply via generative text tools, is precisely the kind of scale efficiency AI critics have warned about since ChatGPT's early days.
OpenAI says it identified the accounts tied to Bad Grammar through its own abuse-detection work and has since banned them, folding the case into a broader pattern the company has been tracking of state-linked actors experimenting with generative AI for influence campaigns rather than building anything technically novel. The report frames this as incremental misuse of existing capability, not some leap in sophistication, and OpenAI is positioning its own detection systems as the primary check against that kind of low-effort, high-volume abuse.
Still, the fact that a foreign influence operation reached for consumer AI tools to write Telegram comments, in two languages, aimed at four separate political fault lines, says something about how normalized this tactic has become. It is no longer a hypothetical risk analysts flag in white papers. It is a line item in a takedown report.
My take — AI-written commentary, not fact-checked reporting
Nobody should be shocked that a state actor pointed a chatbot at Telegram threads instead of hiring a room full of trolls; that was always the cheaper path once these models got good enough to fake fluency in two languages at once. What bugs me is the framing that catching this proves the safety systems work — it proves OpenAI can spot sloppy, low-effort abuse after the fact, which is a much lower bar than preventing the next, better-disguised version. If closed labs want credit for policing misuse, they need to publish more of these reports faster, not drip them out as reputation management dressed up as transparency.
Read more about this at: OpenAI
