OpenAI bans Russia-linked ChatGPT accounts tied to a fake Israeli institute campaign

OpenAI said it banned a cluster of ChatGPT accounts very likely originating in Russia after investigating a covert influence operation centered on the purported Israel-based International Burke Institute, or IBI. The company said ChatGPT was used mainly to produce and distribute promotional material while the operators worked to conceal their Russian origin. 3

The disclosure points to a broader influence tactic: instead of relying only on viral posts, the operation appeared to build a convincing-looking institution, attach borrowed authority to it and distribute its narratives through multiple online communities.

What OpenAI said the operation did

OpenAI said the operators accessed ChatGPT through VPNs because the service is not available in Russia. They prompted mainly in Russian and asked the model to remove linguistic clues that could reveal where they were based. Most of the resulting social posts were written in English. 3

The campaign centered on a website registered in February 2025 that described IBI as an Israel-based expert community. OpenAI said the site contained academic material copied from legitimate work, including material that was misattributed or appeared to have been machine-translated from a Slavic language. 3

The site also promoted a “sovereignty index” that portrayed Russia favorably while denigrating Western countries. Associated messaging criticized Ukraine, the European Union and the German government, and advocated improved relations with Russia. 3

How ChatGPT was used across platforms

According to OpenAI, the accounts generated posts and replies promoting IBI material on Telegram, X, LinkedIn, Facebook and Substack. The activity included multilingual content, channel summaries and visual branding such as logos for Telegram channels targeting audiences in Germany, the United States, France, Poland and Türkiye. 3

One operator generated German-language material for the Telegram channel “Lahme Ente.” OpenAI also described some accounts presented as ordinary users as likely inauthentic, while IBI-branded Substack replies encouraged real users to follow the institute’s channel. 3

The reported use of ChatGPT was therefore not limited to drafting isolated messages. It supported a repeatable content workflow spanning languages, platforms, account identities and branding.

The institute’s borrowed credibility

A key feature of the campaign was the appearance of institutional legitimacy. OpenAI said IBI’s website claimed prominent experts, including Francis Fukuyama and Noam Chomsky, while its associated articles often came from elsewhere online or were connected to incorrect authors. 3

That combination matters because a fabricated or misleading institutional identity can make propaganda appear to be independent analysis. Readers may encounter a polished website, expert names, research-style articles and consistent social branding without seeing the operation behind them.

Limited reach, but a scalable model

OpenAI characterized the campaign’s audience as relatively small and its immediate impact as limited. It assessed the operation at the lower end of Category Three on the Brookings Breakout Scale: activity appeared across multiple platforms, with some evidence of authentic engagement, but without meaningful breakout or broad virality. 3

The company’s concern was the infrastructure rather than the campaign’s current follower count. The operation combined a purported expert institution, copied research, multilingual distribution, coordinated social accounts and efforts to disguise provenance. OpenAI said that structure distinguished it from some earlier Russia-linked operations. 3

The longer-term risk is that a system like this can be reused and scaled. Even modest engagement can provide a foundation for manufacturing apparent authority, directing audiences toward coordinated channels and laundering propaganda through sources that look independent.

The broader lesson

The case illustrates why AI-enabled influence operations are difficult to assess by reach alone. A campaign may not go viral and may still build assets that can be reused: websites, names, logos, translated content, social accounts and a library of apparently authoritative material.

OpenAI’s disclosure does not show that the operation achieved a major breakout. It does show how generative AI can help operators maintain a multi-platform publishing system while concentrating their effort on concealment and credibility. 3