Asger Risom
OpenAI says it banned ChatGPT accounts linked to the pro-Kremlin “Rybar” network after uncovering an AI-assisted influence campaign that generated propaganda, drafted interference proposals in Africa, and seeded coordinated posts across social media.
OpenAI says it has banned a cluster of ChatGPT accounts tied to a likely Russia-origin influence operation that used the AI system to generate propaganda, coordinate messaging, and draft plans for interference efforts abroad.
The campaign, which OpenAI has dubbed “Operation Fish Food,” was linked to the pro-Kremlin “Rybar” network — a Russian-language Telegram brand with a large online following. According to OpenAI, some of the accounts involved in the activity were likely operated from Russia.
The investigation offers a detailed look at how generative AI tools can be used not just to produce individual posts, but to industrialize online influence operations.
AI as a content farm
OpenAI said the banned accounts used ChatGPT to generate social media posts in multiple languages, including Russian, English, and Spanish. Some of that content appeared on “Rybar”-branded accounts and its main website, while other posts were distributed through social media accounts that did not openly disclose any connection to the network.
In addition to written posts, one user generated Sora videos promoting the “Rybar” brand.
The company also identified batches of English-language comments that appeared to be AI-generated and were later posted across X and Telegram accounts. Using open-source investigative techniques, OpenAI said it found exact matches between ChatGPT outputs and comments published by various accounts. Many of those accounts had no publicly declared link to Rybar.
In at least one instance, the operator generated seven tweets from a single prompt. Those tweets were then posted by six different X accounts. According to platform statistics cited in the report, the most-viewed post exceeded 150,000 views, while the least-viewed was seen just 57 times. The account with the largest reach had more than 600,000 followers, compared with 827 followers for the smallest account — suggesting that distribution power depended more on the size of the account than on the AI-generated nature of the content itself.
OpenAI said it could not independently confirm exactly how all of the AI-generated content was published, noting that some of the involved accounts also appeared to post material not generated by its models.
Drafting interference plans
Beyond content generation, OpenAI said the main account also asked ChatGPT to help draft commercial proposals on behalf of “Rybar” for covert interference campaigns in Africa.
According to the report, one prompt requested translation into English of a list of services Rybar could offer to unnamed clients. These allegedly included operating X and Telegram accounts, running a bilingual investigative journalism website focused on Africa, securing paid placements in French-language media, and managing amplification networks.
Another prompt asked the model to edit a proposal for what appeared to be a deployed election interference team in Africa, outlining both online and on-the-ground activities. The plan referenced building local agent networks and organizing large-scale events.
Additional prompts sought information about election processes in Burundi and Cameroon and sketched out campaign options in Madagascar, including the idea of inflaming protests. One proposed project carried an estimated annual budget of up to $600,000.
The same user also generated promotional material for a news outlet called “REST Media,” which open-source researchers have linked to Rybar. In one example cited by OpenAI, the user input a REST Media article accusing Germany of building an influence network in Moldova and then asked ChatGPT to generate comments about it. Several of those comments were later posted to Telegram channels linking back to the article.
Familiar messaging, wide reach
According to OpenAI, the content generated as part of the operation was typical of prior Russian influence campaigns. Posts praised Russia and its allies, criticized Ukraine, and accused Western governments of foreign interference.
The Rybar network has a significant online footprint, with roughly 1.4 million subscribers on its main Russian-language Telegram channel. Many of the X and Telegram accounts that distributed the AI-assisted content also had tens of thousands of followers.
However, OpenAI said it did not observe the campaign’s content being amplified by major mainstream news outlets. Nor did it independently verify on-the-ground activity in Africa that matched the proposals described in the prompts.
Using its internal “Breakout Scale,” which ranks influence operations from 1 (lowest) to 6 (highest), OpenAI assessed the campaign at the top end of Category 3 — indicating spread across multiple communities and platforms, but not a full breakout into broader public discourse.
A window into AI-enabled influence
The case provides a rare look at how generative AI can function as a behind-the-scenes production engine for coordinated influence networks.
Rather than replacing existing tactics, the AI appears to have streamlined them — generating multilingual posts, drafting strategic documents, and producing batches of coordinated social media content at scale.
OpenAI said it banned the accounts involved after detecting the activity.
The episode underscores a growing challenge for tech platforms and governments alike: as generative AI tools become more accessible, influence operations can be planned, refined, and scaled with far less friction than before — even if their ultimate impact still depends heavily on the reach of the accounts pushing the message.
Source: OpenAI report on Covert IO: Operation “Fish Food”
