OpenAI Bans Russian ChatGPT Accounts Behind Fake Think Tank

OpenAI dropped its latest influence-operation takedown report on August 26, and this one is a decent case study in how far a determined actor can stretch a chatbot subscription. The company banned a cluster of ChatGPT accounts operating out of Russia, all routed through VPNs to dodge the regional restrictions OpenAI already has in place. What they built with those accounts is the interesting part.

A think tank that never existed

The operators stood up something called the International Burke Institute, a website registered back in February 2025 that presented itself as an “expert community” based in Israel. It claimed to cover economics, politics, sociology, and international relations, the kind of vague, broad mandate that lets an operation publish about anything while sounding credentialed. The site name-dropped well-known academics, including Francis Fukuyama and Noam Chomsky, as if they were affiliated experts. They were not. OpenAI reviewed a sample of 36 articles tied to the institute’s listed experts and found 34 were lifted from elsewhere on the internet, some years old, several attributed to the wrong authors entirely, and a number that read like machine translations from Slavic-language originals.

That’s the tell here. The institute wasn’t producing original analysis so much as repackaging scraped content under invented bylines, then using ChatGPT to smooth the English, write social copy promoting the “research,” and even generate profile pictures for the Telegram channels distributing it.

The sovereignty index angle

The most calculated piece of the operation was a scoring system the institute called its sovereignty index, ranking countries across seven dimensions: political, economic, technological, information, cultural, cognitive, and military sovereignty. Russia landed at 601.4 points, the United States at 650.9. Close enough to look like a defensible, data-driven ranking rather than propaganda, while still nudging the reader toward parity between the two. That’s the kind of framing a human propagandist would spend weeks polishing. Here it looks like it was assembled and iterated on quickly, with the model doing a lot of the drafting and editing work.

Content built around the index and the institute’s other output got pushed across Substack, Telegram, X, Facebook, and LinkedIn, largely in an attempt to inject pro-Kremlin talking points into discussions where a Western-sourced comment would carry less weight than one from an ostensibly neutral, Israel-based analyst.

Reach was small, the pattern isn’t

To OpenAI’s credit, they’re upfront that this campaign didn’t move the needle much. Telegram channels tied to the operation topped out around 10,000 to 20,000 followers, and engagement on the other platforms was thin. Nobody should read this as evidence that AI-assisted influence ops are suddenly effective at scale. What’s notable is the production model: a small team used a commercial chatbot to fake an institution’s entire content pipeline, from research summaries down to the visual branding, in a fraction of the time and headcount that kind of front used to require.

This is the fourth or fifth report of its kind OpenAI has put out since it started publishing periodic takedowns of misuse on its platform, and the throughline across all of them is consistent. State-linked actors aren’t using AI to invent radically new tactics so much as to run existing disinformation playbooks faster and cheaper. Fake personas, plagiarized “research,” astroturfed social presence: all standard influence-op ingredients, just assembled with a lot less manual labor than before.

Why it matters for defenders

For security teams and platform trust and safety units, the actionable signal isn’t the specific Burke Institute domain, which is already dead weight now that it’s been outed. It’s the pattern: watch for suspiciously prolific “expert” sites with thin verifiable histories, scoring systems or indices that conveniently favor a particular geopolitical narrative, and content clusters that show up simultaneously across five or six platforms with near-identical framing. Those are increasingly signs of AI-assisted production rather than a lone blogger with strong opinions, and detection tooling built around content provenance and cross-platform correlation is going to matter more as this keeps happening.

References

• The Hacker News — OpenAI Bans Russian ChatGPT Accounts Used to Run Influence Operation
• OpenAI — Disrupting a new covert influence campaign from Russia
• TechRepublic — OpenAI Bans Russian ChatGPT Accounts Used in Covert Influence Campaign