OpenAI says it has banned a cluster of ChatGPT accounts likely originating in Russia after discovering their role in a covert influence campaign built around a fictitious think tank. The operation used AI-generated social media posts, fake identities and a Russia-friendly “sovereignty” index to distribute political narratives across Western platforms.

The campaign promoted the International Burke Institute (IBI), which presented itself as an Israel-based expert community. OpenAI’s investigation found that the organization’s website contained copied and misattributed academic material, while associated accounts posted content on X, LinkedIn, Facebook, Substack and Telegram. The company said the operation’s observed audience was limited, but its infrastructure appeared designed for wider distribution.

What OpenAI found

According to OpenAI, the operators accessed ChatGPT from Russia using VPNs because the service is not available there. They prompted in Russian to produce mostly English-language social media comments and instructed the model to remove linguistic clues that might reveal their origin.

The main objective was to promote IBI articles and its “sovereignty index,” which ranked Russia favorably while portraying Western countries negatively. The campaign’s posts appeared through accounts carrying IBI branding as well as accounts that looked like ordinary users. Some operators also generated replies to real Substack users, using those interactions to encourage people to follow IBI channels.

The operation extended beyond English-language promotion. OpenAI identified German-language posts on a Telegram channel called “Lahme Ente,” or “Lame Duck,” which criticized Ukraine, the European Union and Germany’s federal government while advocating closer relations with Russia.

A second operator generated logos for roughly a dozen Telegram channels focused on Germany, the United States, France, Poland and Türkiye. That operator also asked ChatGPT for Russian-language summaries of the channels’ activity. OpenAI said some of the channels occasionally promoted IBI content, while at least one presented itself as an American outlet despite showing signs of non-native English.

The evidence and its limits

The strongest evidence in this case comes from OpenAI’s own investigation and account-enforcement action. The company describes the accounts as “very likely” originating in Russia, rather than presenting a public legal attribution to a specific state agency or named organization. That distinction matters: the evidence supports a Russia-linked operation, but the available material does not establish who ultimately directed or financed it.

OpenAI also reported that 34 of a sample of 36 IBI articles linked to supposed experts and published between September 2025 and May 2026 had been copied from elsewhere online. Some were given false author credits. One article apparently drawn from Cambridge University Press material was attributed to a University of Nottingham professor, while a migration article from the Migration Policy Institute was credited to an Australian professor of food chemistry.

The company said some IBI material appeared to have been written by a Slavic-language speaker and machine translated. OpenAI pointed to the use of “Svetofor,” a transliteration of the Russian word for traffic light, in a reference to Germany’s “traffic-light” coalition. Such clues helped connect the operation’s public-facing material to a Russian-language origin, although they are indicators rather than conclusive proof of individual authorship.

OpenAI characterized the campaign as previously unreported and more elaborate than other Russia-linked operations it has disrupted. The Decoder, which reviewed the company’s findings, reported that individual posts received few views and IBI’s official accounts had small subscriber counts. The linked Telegram channels were larger, with reported audiences of 10,000 to 20,000 followers each. Those figures come from the reporting and OpenAI’s assessment, not from an independent measurement of active or authentic users.

OpenAI placed the campaign at level three on the six-category Brookings Breakout Scale, meaning it had spread across multiple platforms with early indications of reaching real users. That is an assessment of campaign development, not a measure of persuasive impact or electoral influence.

Where ChatGPT fit in the operation

The campaign’s use of ChatGPT was operational rather than centered on producing the IBI website’s research. OpenAI said the model generated social media posts, replies, summaries and some visual assets such as channel logos. The website’s copied articles and “sovereignty” reports were not generated by OpenAI models, based on the company’s review.

That division is important for builders and platform operators. Generative AI did not need to create a complete propaganda publication for the operation to benefit from automation. It helped produce variations of messages, adapt content for different languages and platforms, and maintain a network of accounts and channels. The effect is to reduce the labor required for routine distribution, even when the underlying claims and source material are fraudulent.

OpenAI said it previously identified a Russian network called “Bad Grammar” in 2024, which used its models to generate political comments on Telegram about Russia, Ukraine and the Baltic states. The company also reported “Operation Helgoland Bite,” a later campaign that used ChatGPT to produce German-language material ahead of Germany’s 2025 federal election.

These cases only cover activity visible to OpenAI. The Decoder noted that operators can also use other providers and open-weight models, making the company’s enforcement data an incomplete view of AI-assisted influence activity. OpenAI’s account bans therefore demonstrate disruption on one service, not containment of the broader tactic.

Implications for AI platforms and enterprises

For AI companies, the incident shows why abuse monitoring must examine behavior across prompts, accounts and external platforms rather than treating each generation request in isolation. Requests to translate political messaging, summarize channel activity or remove linguistic markers may appear benign individually. In combination, they can reveal coordinated deception.

For social platforms, the case reinforces the need to evaluate distribution networks, not only the factual accuracy of individual posts. A message generated by ChatGPT may be difficult to classify as harmful on its own. The stronger signals may be repeated promotion of a new organization, synchronized posting, fabricated biographies, copied research and accounts whose activity is almost entirely amplification.

Enterprise AI teams should draw a narrower but practical lesson: provenance and identity controls matter when generative tools are connected to publishing workflows. Logging, rate limits, review for political or reputational content and detection of repeated cross-platform patterns can reduce the chance that internal automation is repurposed for covert outreach. Those controls will not determine the operator’s intent by themselves, but they can make abuse more visible.

What to watch next

The immediate signal will be whether OpenAI or other platforms identify additional accounts, domains or channels connected to the International Burke Institute. Researchers should also examine whether the Telegram audiences represent genuine followers, coordinated accounts or dormant infrastructure.

A second question is whether similar operations are shifting toward open-weight models or services with weaker abuse reporting. That would test whether platform-level bans meaningfully slow campaigns or simply push operators to different tools.

Finally, investigators will be watching for more sophisticated localization: natural-sounding translations, credible author profiles and content tailored to specific political communities. The current operation showed visible weaknesses, but OpenAI’s warning is that its infrastructure could have been scaled before those weaknesses were corrected.

Creati.ai perspective

This incident is less about ChatGPT independently creating influence than about operators embedding a general-purpose model inside a broader deception system. The fake institution, stolen research, fabricated identities and distribution channels supplied the campaign’s credibility and reach; AI reduced the cost of producing and adapting the supporting content.

For the AI industry, the durable challenge is attribution and coordination. Account bans can remove one campaign from one platform, but effective defense will require cooperation among model providers, social networks, researchers and investigators while preserving clear boundaries around evidence. OpenAI’s report is significant because it exposes that workflow in detail, but its own findings also show why a single provider cannot measure the full scale of AI-assisted influence operations.