Que s'est-il passé
OpenAI says it disrupted a previously unreported covert influence campaign that very likely originated in Russia. Operators used ChatGPT to generate social-media posts and replies promoting the International Burke Institute, a self-described expert community that presented itself as being based in Israel. OpenAI says the campaign used VPNs because its services are not available from Russia.
OpenAI says it recently banned a cluster of ChatGPT accounts that very likely originated in Russia. The operators prompted in Russian to generate comments and posts, many of them in English, for Substack, Telegram, X, Facebook and LinkedIn. They instructed ChatGPT to conceal linguistic clues that could reveal a Russian origin, and used VPNs to access the service because OpenAI says it does not allow access from Russia. The generated material primarily promoted the International Burke Institute, or IBI, and its website.
The campaign’s AI activity was focused on social distribution rather than on producing the institute’s main website content. OpenAI identified posts promoting IBI articles, including posts from accounts using the institute’s name and logo and from accounts that appeared to be ordinary users but mainly shared IBI material. Operators also generated replies to real Substack users that promoted the IBI channel. Other activity included German-language Telegram posts criticizing Ukraine, the European Union and the German government, as well as posts advocating better relations with Russia. One operator generated logos for about a dozen Telegram channels focused on Germany, the United States, France, Poland and Türkiye, while another requested Russian-language summaries of channel activity.
OpenAI describes the IBI website as the apparent core of the operation. The site was registered in February 2025 and claimed to be based in Israel. In a review of 36 articles linked to named experts and published between September 2025 and May 2026, OpenAI says 34 were copied from elsewhere online. Some were years old or attributed to the wrong authors. The source gives examples involving material originally published by Cambridge University Press and the Migration Policy Institute. The site also published country reports built around a sovereignty index that praised Russia and criticized countries including France, Germany, the United States and the European Union.
The source says the campaign’s most notable was the combination of AI-generated outreach with a fabricated or misleading authority structure. The website used purported experts, republished academic work and a proprietary-looking index to make its narratives appear more credible. OpenAI says some online traces suggest that a small number of real individuals in Israel may have submitted articles or represented IBI at conferences, but it cannot determine their relationship with the Russian operators or with the wider campaign.
Détails de la source: openai.com ↗
Pourquoi c'est important
The operation shows how can support a broader influence infrastructure rather than act as the entire campaign. The account cluster used ChatGPT to produce promotional material, while the campaign relied on a plausible-looking website, copied academic work, purported experts and a sovereignty index designed to favor Russia and criticize Western countries.
The case is significant because ChatGPT was used as one component of a larger effort to manufacture authority. The operators did not need the model to create every part of the operation. They used it to produce repeated, adaptable social-media content that directed audiences toward an institution designed to look independent and expert-led. That division of labor can make influence activity harder to identify if investigators look only for fully AI-generated websites or obvious automated accounts.
OpenAI says the operation’s immediate audience appears to have been limited. Typical social posts received low numbers of views, and official IBI accounts had low subscriber counts. The Telegram channels were more substantial, generally reporting 10,000 to 20,000 followers each. Using the Brookings Breakout Scale, OpenAI assessed the operation at the lower end of Category Three, meaning activity across multiple platforms with some indications of breakout to authentic audiences. Those figures are the company’s assessment, not an independently verified measurement presented in the source.
The broader practical risk lies in the infrastructure created for future use. A credible-looking site, a network of topic-specific channels, attributed articles and a signature index can remain available even when individual posts have little reach. Such assets could be reused, expanded or redirected toward new narratives. The source does not show that this campaign achieved broad persuasion or changed political behavior, so its significance should not be overstated. It does show a concrete method for combining synthetic content production with deceptive identity-building.
The case also illustrates a detection tradeoff. OpenAI says its investigation began with AI-generated social-media posts but expanded after those posts led investigators to the IBI website and its copied or misattributed material. In that account, AI use helped expose the operation because the generated content created observable links among accounts, channels and the institute. The source does not establish whether comparable investigations by other platforms or researchers reached the same findings.
Mécanisme interactif : comment cela fonctionne réellement
Explorez de manière interactive la technologie sous-jacente à ce développement.
crm_get_transaction(id='4092').Why can ethical evaluation not be reduced to one model score?
Que regarder ensuite
The source leaves important questions unresolved, including the identities of the operators, their relationship with people in Israel who may have represented the institute, the campaign’s total reach and whether other AI systems were used. Further monitoring should focus on whether the institute’s assets, social accounts or related channels reappear under new names.
The first priority is whether the disrupted network returns through replacement accounts, domains or channels. OpenAI says it banned a cluster of accounts, but the source does not say that every associated social account, website asset or Telegram channel was removed. Monitoring should therefore track recurring IBI references, the sovereignty index, copied articles, distinctive language patterns and channel networks spanning the countries identified in the report.
The relationship between the Russian operators and the apparent Israeli participants remains a central unknown. OpenAI says some online traces suggest that real individuals in Israel may have submitted articles or appeared at conferences, but it cannot determine whether they knowingly participated, were misrepresented or had any operational connection to the Russian accounts. That distinction matters for accurately describing responsibility and for avoiding unsupported conclusions about the people or organizations named on the site.
The source also does not quantify how many posts were generated, how long the ChatGPT accounts were active, how many authentic users encountered the material or whether the operators used other AI services. It does not report any confirmed policy response from the social platforms where the posts appeared. Those gaps limit what can be concluded about the operation’s scale, coordination and effectiveness.
For researchers and platforms, the case points to the value of examining the links between AI use and non-AI assets. Useful indicators may include repeated multilingual generation, efforts to suppress clues about an operator’s origin, clusters of accounts promoting the same purported institution, and websites that combine copied scholarship with invented authority signals. The source establishes these features in this case, but it does not establish that they form a universal detection rule or that every similar pattern reflects a state-linked operation.