OpenAI Says It Dismantled Russian Network That Built a Fake Think Tank With ChatGPT

OpenAI said Tuesday that it had identified and shut down a cluster of ChatGPT accounts originating in Russia that were used to build and promote a fabricated Israeli think tank as part of a covert online influence campaign, in a report published on the company's site describing the operation in detail. The network used the chatbot to generate articles, social media posts and institutional branding for an outfit calling itself the International Burke Institute, which presented itself as an independent "expert community" while pushing narratives favorable to the Kremlin.

The company said the operators routed their access through virtual private networks to get around Russia's exclusion from ChatGPT, then wrote prompts in Russian while instructing the model to produce finished text in English and German. According to OpenAI, the operators explicitly told the chatbot to strip out any linguistic patterns that might betray Russian authorship, an instruction the company said points to a deliberate effort at concealment rather than incidental misuse.

How the operation worked

The fabricated institute's website was registered in February 2025 under a claimed Israeli street address. It listed a roster of purported affiliated experts and, according to OpenAI, falsely associated the site with well-known academics including the political scientist Francis Fukuyama and the linguist Noam Chomsky, neither of whom appears to have any actual connection to the project. In a review of 36 articles attributed to the institute's contributors and published between September 2025 and May 2026, OpenAI found that 34 had been lifted from other sources online, in some cases years old, often with authorship altered or misattributed.

A centerpiece of the campaign was what the operators called a "sovereignty index," a country-by-country scoring system designed to portray Russia favorably while criticizing Western governments and institutions. Content built around the index, along with other material generated by the network, was distributed across five platforms.

• 36 articles reviewed from the fabricated institute's site; 34 were plagiarized or misattributed
• Website registered in February 2025, styled as an Israeli research organization
• Roughly a dozen Telegram channels received AI-generated logos, targeting Germany, the United States, France, Poland and Türkiye
• Those channels drew between 10,000 and 20,000 followers apiece

OpenAI measured the campaign's spread using the Breakout Scale, a framework developed by the Brookings Institution to gauge how far influence-operation content travels beyond a network's own accounts. The company scored the effort's Substack presence at 10 out of 100 and its Telegram activity at 57 out of 100, placing the overall campaign at the low end of "Category Three" — meaning the content reached multiple platforms but achieved only limited traction with real audiences. One thread of German-language Telegram posts, aimed at criticizing Ukraine, the European Union and Germany's government while promoting closer ties with Moscow, contained a telling error: it referenced Germany's "traffic-light coalition" using "svetofor," a Slavic word for traffic light, rather than the correct German term — a slip OpenAI cited as evidence of the content's non-native origin.

Part of a recurring disclosure

Tuesday's report is the latest in a series of periodic disclosures OpenAI has published since February 2024 documenting attempts to misuse its models for deceptive or malicious purposes, cataloged on the company's running index of threat-intelligence findings. In its most recent prior installment, published in October, the company said it had disrupted more than 40 networks that violated its usage policies since it began the reporting effort. Earlier this year, OpenAI disclosed a separate Russian-linked network that had drafted budget plans of up to $600,000 for an influence campaign targeting several African countries, underscoring that state-linked actors have repeatedly tested the platform's guardrails rather than been deterred by earlier takedowns.

The pattern reflects a broader dynamic facing generative-AI companies: the same tools that let ordinary users draft essays or summarize documents also let influence operators produce large volumes of polished, multilingual content at low cost, then disguise its origin. OpenAI's disclosures have become a recurring data point for researchers and journalists tracking how state-linked actors are adapting propaganda tactics to commercial AI systems, a shift that outside coverage of the report, including from wire reporting carried Tuesday, framed as notable less for its reach than for the sophistication of the fabricated institutional front it built.

Limited reach, larger warning

OpenAI was explicit that the campaign's practical impact was modest. Social media posts tied to the network drew low view counts, and the fabricated institute's own accounts had few followers. But the company argued the case was still significant because of how much infrastructure the operators built around a small audience.

The operation illustrates "how AI tools can help bad actors create an appearance of institutional credibility," OpenAI said in its report.

The company also noted that, unlike simpler propaganda efforts it has previously disrupted, this operation's layered construction — a branded institute, a fabricated expert roster, a scoring index and coordinated multilingual social accounts — set it apart from other Russia-linked networks it has taken down, even though it reached a comparatively small audience, according to a summary of the findings published by the Ukrainian outlet dev.ua.

Those most directly affected fall into two groups: the audiences in Germany, the United States, France, Poland and Türkiye that the Telegram channels specifically targeted, and the platforms — Substack, Telegram, X, Facebook and LinkedIn — whose systems hosted the fabricated content without, in OpenAI's account, any indication that the operators' underlying identity had been detected before ChatGPT access was cut off. It was not immediately clear whether any of those platforms had independently removed the accounts and channels described in OpenAI's report, and none had issued a public statement on the findings as of Tuesday afternoon. Fukuyama, Chomsky and representatives for the named platforms did not appear to have publicly addressed the disclosure either, according to coverage of the takedown.

What happens next

OpenAI said it banned the accounts associated with the network and indicated it would continue sharing indicators about the activity with other companies and researchers to help limit further distribution, consistent with the practice it has described in earlier reports. The company has not said whether it expects further enforcement action beyond banning the accounts identified so far, and it did not name the individuals or organization believed to be behind the operation.

For now, the disclosure adds another documented case to a growing public record of how state-linked operators are experimenting with commercial AI tools to manufacture the appearance of independent expertise, even when the resulting content fails to reach a wide audience. With OpenAI continuing to publish these findings on a recurring basis, researchers and rival platforms are likely to keep watching for whether similar fabricated-institute tactics resurface elsewhere, and whether the five social platforms named in Tuesday's report take their own action against the accounts and channels OpenAI identified.