OpenAI’s Russia Takedown Turns a Fake Israeli Think Tank Into an AI Forensics Case Study

OpenAI says it has banned a cluster of ChatGPT accounts that very likely originated in Russia, after tracing them to a covert influence operation built around a fake “expert community” called the International Burke Institute (IBI). IBI presented itself as an independent Israel-based think tank publishing a “sovereignty index” that praised Russia and criticized Western governments. According to OpenAI, it is the most elaborate Russia-linked influence operation the company has disrupted since the war in Ukraine began, and the case is notable less for its reach, which OpenAI describes as small, than for how it was built and how it eventually gave itself away.

A Think Tank Built to Look Independent

IBI’s website was registered in February 2025 and claimed the institute was based in Israel. OpenAI’s investigation found that the articles published on the site were not themselves generated by its models. Many were copied from real academic writing, sometimes with false attribution, while others appear to have been drafted by a Slavic speaker and then machine translated into English. The centerpiece of the operation was IBI’s own “sovereignty index,” a framework the front used to cast Russia in a favorable light while criticizing Western countries.

That distinction matters for understanding what generative AI actually contributed here. OpenAI is explicit that the site’s articles were not produced by its models: the foundational content, the plagiarized essays and the index itself, was copied, translated or drafted through other means. ChatGPT’s documented role was almost entirely downstream: promoting a front that had already been built, rather than authoring it from scratch.

What ChatGPT Actually Did

Amplifying the front

OpenAI says the operation’s main use of ChatGPT was generating social media posts that promoted articles on the IBI website. Investigators identified these posts on X, LinkedIn, Facebook, Substack and Telegram. Some carried the IBI name and logo directly; others came from accounts posing as ordinary users whose main activity consisted of sharing IBI articles. The operators also generated replies to posts from real Substack users on unrelated topics, typically appending a request that the reader follow the IBI channel.

A wider web of channels

Alongside the IBI-focused activity, one operator generated German-language posts for a Telegram channel called “Lahme Ente” (“lame duck”), which routinely criticized Ukraine, the European Union and the German government while advocating closer ties with Russia. A second operator generated logos for a dozen Telegram channels targeting audiences in Germany, the United States, France, Poland and Türkiye, some of which occasionally posted about or shared IBI’s own content. That operator also repeatedly asked ChatGPT for Russian-language summaries of the channels’ activity.

One of the US-focused channels, styled as an American outlet called “American Observer,” had a bio with clear signs of non-native English, including a phrase OpenAI’s report highlights as especially telling: “a totally unhackneyed perspective on hazzy.” The same operator generated a profile picture for the channel.

Covering its tracks

The operators prompted ChatGPT in Russian to produce social media comments, most of which came back in English, and explicitly instructed the model to strip out any linguistic clues that would identify them as Russian speakers. Because OpenAI does not support ChatGPT access from Russia, the operators relied on VPNs to reach the platform at all.

The Word That Undid It

Despite that effort, the operation left a structural tell that survived translation and editing. An IBI article about Germany referred to “the Svetofor coalition,” an odd phrase to describe what OpenAI’s report calls Germany’s coalition of socialists, liberals and greens, known in German as the “Ampelkoalition,” or traffic-light coalition, named for the parties’ red, yellow and green colors. “Svetofor” (светофор) is simply the word for traffic light in Russian and several other Slavic languages. No English or German speaker writing independently would reach for a Russian loanword to describe a domestic German political alliance; OpenAI concluded the phrase was almost certainly a literal machine translation from a Slavic-language original.

It is a small detail, but it is exactly the kind of artifact that content laundering struggles to erase. An operation can copy real academic writing, register a plausible-sounding institute, and instruct a model to hide obvious linguistic markers, yet still leave behind a single mistranslated idiom that only makes sense if the underlying text passed through a Slavic language first. That gap between what a front claims to be and how its content was actually produced is where this kind of forensic work happens.

A Small Audience, an Unusual Amount of Effort

OpenAI is direct about the operation’s limited impact: “the campaign appears to have reached relatively small audiences.” What sets it apart, in the company’s own words, is that “its elaborate construction distinguishes it from other Russia-linked influence operations we have disrupted since the start of the war in Ukraine,” and that this is “the first time we have disrupted an influence operation tied to Russia that went to such elaborate lengths.”

IBI’s website was still online at the time of writing, describing itself in much the same terms OpenAI’s report lays out, a “community of leading experts” offering “thought leadership” on sovereignty. Its own about page carries the same rough-edged English the report describes elsewhere: it dates its mission to “the 21th century,” and its footer copyright notice refers to itself as “The Internationalal Burke Institute.” Small typos like these are exactly the kind of signal that separates a hastily assembled front from the polished, professionally maintained institutions it imitates.

Part of a Bigger Pattern

This is not OpenAI’s first disclosure of Russia-linked influence activity. The company points to a series of similar operations it has disrupted since the war in Ukraine began, of which this is described as the most elaborate to date. The report lands the same day Alabama’s attorney general subpoenaed OpenAI over a separate matter, its Hugging Face account breach, part of a broader stretch of outside scrutiny on the company’s safety and security posture.

The bigger pattern is worth sitting with. Generative AI did not create this operation’s ambitions, the plagiarized essays and the invented “sovereignty index” were built by hand, but it did lower the cost of running the amplification layer: drafting posts across five platforms, generating replies to real users, building logos for a dozen Telegram channels, and summarizing chatter back to the operators in their own language. At the same time, that same AI-assisted production process left behind exactly the kind of structural artifact, a phrase that only makes sense if it passed through Russian first, that trained investigators know to look for. The arms race between AI-assisted influence operations and AI-assisted detection is not new, but this case is a clean illustration of how it actually plays out in practice: more content, produced faster, still carrying tells that give it away.