OpenAI exposes Russian influence campaign that promoted a fictitious Israeli "research institute"
OpenAI has exposed a Russian influence campaign that exploited AI to promote a fictitious Israeli "research institute" called the International Burke Institute (IBI). The operation used plagiarized academic work and fabricated indices to present Russia in a positive light.
OpenAI has exposed a Russian influence campaign that exploited its AI capabilities to promote a fictitious Israeli "research institute" called the International Burke Institute (IBI).
"What began as an investigation into social media posts generated using AI led to the exposure of a much broader influence campaign, built around a website that contained plagiarized academic work, a 'sovereignty index' that presented Russia in a positive light, and efforts to disguise the Russian origin of the activity," the company stated in a report published yesterday.
The influence campaign was exposed after OpenAI identified and blocked a cluster of ChatGPT accounts whose operator wrote prompts in Russian to generate posts and comments on Substack, Telegram, X, Facebook, and LinkedIn. "Most of the responses were in English, and the operator sought to disguise linguistic markers of the Russian origin. We do not allow access to our models from Russia, and a VPN was used to access our platform," the company stated.
Most of the generated content sought to promote the fictitious research institute IBI. The institute's website claims it is based in Israel, and many of the articles published on it were actually copied from real academic publications, with incorrect attribution. Others were written in a Slavic language and automatically translated into English. The most prominent publication on the institute's website was a "sovereignty index" intended to praise Russia and harm Western countries.
The primary use of ChatGPT was for creating posts that promoted the institute. "Some were published by accounts that included the name and logo of IBI, others appeared to come from regular users, whose main activity was posting IBI posts," OpenAI stated. "In addition, the operators generated responses to posts by real users on Substack on a variety of topics. The responses often included a request to follow the IBI channel on the platform and were posted by it."
Additionally, one of the operators generated posts in German that were published on the Telegram channel Lahme Ente (Lame Duck), which criticized Ukraine, the European Union, and the German government. Another operator generated logos for a dozen Telegram channels, which focused on Germany, the USA, France, Poland, and Turkey.
The AI activity identified by OpenAI was only complementary to the core work of the influence campaign, which was operating the fictitious research institute. The IBI website claimed it was an independent "community of experts" operating from Israel, which included world-renowned experts such as Francis Fukuyama and Noam Chomsky. But a sample examination of 36 articles published on the site, conducted by the company, revealed that 34 of them were copied from other places.
"Some were several years old, others were attributed to incorrect authors. For example, one article about the China-Pakistan economic corridor was copied from a publication by Cambridge University Press and incorrectly attributed to a professor from the University of Nottingham. This behavior indicates a desire to give IBI a more legitimate image by incorporating authentic content."
Alongside the articles, the IBI website includes reports on the sovereignty of various countries. Most of the reports, which were not generated by AI, included criticism of Western countries. According to OpenAI, there is online evidence that individuals in Israel mentioned or represented IBI in articles they submitted for publication and at conferences they attended. "We cannot determine what the relationship is between these individuals, IBI, and the operators in Russia," the company said.
Despite this, OpenAI estimates that the impact of the campaign was limited: the posts that were published recorded a low number of views, and the IBI accounts had a small number of subscribers. The campaign's Telegram channels recorded between 10,000 and 20,000 followers.
"The importance of the campaign lies less in the audience it reached, and more in the infrastructure it built," it was stated. "While the operators used ChatGPT only to generate isolated posts, these posts directed to a reliable-looking institute, with alleged experts, republished academic works, and an allegedly unique index. This illustrates how influence actors can use AI as a supporting tool as part of a broader effort to generate authority, blur the origin of preferred narratives, and establish assets that can be expanded in scope over time."
