Fake Think Tank, Real AI: Inside OpenAI's Takedown of Russia's 'Burke Institute' Influence Operation

On August 25, 2026, OpenAI published a report that reads less like a takedown notice and more like a case study in modern information warfare. The company banned a cluster of ChatGPT accounts that, in its assessment, “very likely originated in Russia.” Their mission was not to hack anything, generate deepfakes, or flood platforms with bot armies. It was something quieter and arguably more insidious: to manufacture credibility — a whole fake institution of it — and then use AI-generated social media chatter to make that institution look like it had an audience.

The operation centered on the International Burke Institute (IBI), a self-described “expert community” that claimed to be based in Israel. Its website, registered in February 2025, presented a polished facade: expert profiles invoking famous names like Francis Fukuyama and Noam Chomsky, a library of academic-looking articles, country reports, and a proprietary index that ranked nations on their “sovereignty.” To a casual reader — or an algorithmic feed — it looked like a legitimate policy shop. It was neither legitimate nor, in any meaningful sense, a shop.

What the operation actually built

OpenAI’s investigation began, as these things now do, with anomalies in how its own model was being used. Analysts noticed ChatGPT accounts prompting in Russian to generate social media comments and posts, mostly in English, targeting X, Facebook, LinkedIn, Telegram, and Substack. The operators gave the model a specific and revealing instruction: hide any linguistic clues that the writers were Russian. Because OpenAI does not allow access from Russia, the operators tunneled in through VPNs.

“What began as an investigation into AI-generated social media posts led us to a much broader influence operation,” OpenAI wrote, “built around a website containing copied and misattributed academic work, a ‘sovereignty’ index that cast Russia in a favourable light, and efforts to disguise the operators’ Russian origins.”

The website’s academic facade was largely hollow. OpenAI’s review found that 34 of 36 sampled articles published under expert profiles had been copied from elsewhere online — some years old, others attributed to the wrong authors entirely. One article on the China-Pakistan Economic Corridor appears to have been lifted from a Cambridge University Press original and misattributed to a University of Nottingham professor whose actual expertise is South Asian politics. Notably, OpenAI found no evidence the site’s articles themselves were written by ChatGPT; some material seemed to have been drafted by a Slavic-language speaker and machine-translated. The AI’s role was narrower but essential: it built the promotional layer — the posts, comments, and channel branding — that made the operation look organic.

The operation’s signature artifact was the “Sovereignty Index,” also called the Burke Index. It scored countries across seven dimensions — political, economic, technological, information, cultural, cognitive, and military sovereignty — using “equalization coefficients” and “expert assessments” that no outsider could audit. The results were calibrated for narrative effect: consistent praise for Russia’s “record military and economic self-sufficiency” and “full digital and military nuclear sovereignty,” paired with criticism of Western states, especially France, Germany, the EU, and the United States. A made-up index is effective precisely because numbers carry authority even when the methodology is vague, selective, or fabricated.

Reach versus infrastructure

By classic influence-operation metrics, the campaign underperformed. OpenAI assessed that it reached “relatively small audiences,” with the associated Telegram channels — including a German-language channel tellingly named “Lahme Ente” (“lame duck”) that attacked Ukraine, the EU, and the German government — each amassing roughly 10,000 to 20,000 followers. One operator used ChatGPT to generate logos and profile pictures for a dozen country-focused channels covering Germany, France, Poland, Türkiye, and the United States, then repeatedly requested Russian-language summaries of their activity — a telling operational habit that gave OpenAI insight into the network’s own command structure.

But OpenAI’s report draws the opposite lesson from the numbers: “The significance of the operation lies less in the audience it reached, however, than in the infrastructure it had built.” The actors had constructed an entire counterfeit institution — purported experts, republished academic work, a proprietary risk index — an asset that could, in principle, be scaled indefinitely. The cheap AI-generated promotion was just the seed money.

There is a second lesson embedded in the story, and OpenAI states it plainly: “It also illustrates how their supporting use of AI can lead to the broader operation being exposed.” The very tool that lowered the cost of manufacturing an audience also left the fingerprints that unraveled the network. Operators who prompted in Russian while producing English content, who requested summaries of their own channel activity, and who needed logos generated for a fleet of country-branded channels created a behavioral signature that platform-level visibility caught.

Context: a shifting doctrine

The Burke Institute case sits inside a broader trend that security researchers have tracked since the start of the war in Ukraine: influence operations are migrating from volume-based strategies — bot armies, mass-posting, coordinated inauthentic behavior at scale — toward credibility-based strategies. The new model does not try to win arguments; it tries to build institutions that appear entitled to win them. A think tank with an index is quotable. A citation from “the International Burke Institute” smuggles a narrative into op-eds, cable news chyrons, and eventually policy debates, laundered through an academic-sounding veneer.

This is also why the operation’s Israel cover story matters. Masking a Russian influence asset as a neutral, third-country expert community is designed to defeat the simplest heuristic most readers apply: “who is saying this?” The answer the campaign wanted readers to reach was “independent experts in Israel.” The truthful answer, per OpenAI’s assessment, was operators in Russia hiding behind VPNs.

For AI labs, the incident reinforces a doctrine that has been hardening across the industry: the model provider is now a detection surface. OpenAI has disrupted more than a dozen influence operations since the war in Ukraine began, and each takedown report doubles as a signal to adversaries about exactly what telemetry gets an operation caught. The uncomfortable corollary is that the next iteration will be built with that feedback in mind — prompting through cleaner proxies, splitting content generation across providers, and investing even more in the human-plausible institutional facade that AI no longer needs to supply.

What to watch

Three things make this case worth remembering. First, the counterfeit-institution playbook now has a documented template: register a plausible domain, populate it with plagiarized scholarship, invent an index, then use cheap AI promotion to simulate organic traction. Expect variations of this aimed at other regions — Italian and European audiences were already on the campaign’s target list, and the multi-country channel fleet suggests ambitions well beyond Germany.

Second, attribution increasingly comes from usage patterns, not content. Nothing the operation published was detectably “AI-written” in a way that mattered; what exposed it was how the operators behaved inside the tool. That shifts detection from forensics on the message to forensics on the messenger’s workflow.

Third, the gap between infrastructure and reach is a warning about latency. The Burke network was caught early, at small scale. The failure mode to fear is not this operation, but the one that scales its institutional facade for a year before anyone looks — and by the time anyone looks, the index is being cited in earnest.

OpenAI’s ban removed the campaign’s AI scaffolding, but the website, the index, and the Telegram channels largely persist. The infrastructure built to manufacture authority is still standing. The only thing that changed is that we now know exactly how it was made.