OpenAI Bans Russian ChatGPT Accounts Used in Covert Influence Operation
OpenAI has banned a cluster of ChatGPT accounts assessed to have originated in Russia after investigators linked them to a covert influence operation promoting the International Burke Institute. The operators used VPNs to bypass regional access restrictions and generated mostly English-language social media content designed to conceal Russian linguistic traces and promote narratives favorable to Russia.
OpenAI has disrupted a cluster of ChatGPT accounts that it says were used as part of a Russia-linked covert influence operation targeting audiences across several major social media platforms.
According to OpenAI, the accounts were very likely operated from Russia and used VPN services to access ChatGPT despite OpenAI not offering access to its models from Russia.
The operators used the AI system primarily to generate social media posts and comments published across:
- Substack
- Telegram
- X
Most of the generated content was written in English, while the operators explicitly instructed ChatGPT to avoid language patterns that might reveal their Russian origin.
International Burke Institute at the Center of the Campaign
Much of the content promoted an organization called the International Burke Institute (IBI).
The group presented itself as an Israel-based “expert community” covering topics including:
- Economics
- Politics
- Sociology
- International relations
OpenAI's investigation found that the operation extended far beyond isolated social media comments.
The IBI website contained what OpenAI described as:
- Copied academic work
- Misattributed research
- Purported expert profiles
- Country reports
- A proprietary “Sovereignty Index”
The broader effect was to create the appearance of an established research institution capable of lending credibility to political narratives.
No Evidence That ChatGPT Wrote the IBI Articles
An important distinction in OpenAI's findings is that investigators did not find evidence that the articles published on the IBI website were generated with ChatGPT.
Instead, many of the articles appeared to have been copied from existing academic publications, sometimes with incorrect or misleading attribution.
Other content appeared to have been written by Slavic-language speakers and machine-translated into English.
ChatGPT's observed role was more limited and focused primarily on producing promotional social media material around the broader influence infrastructure.
Operators Tried to Hide Their Russian Origin
OpenAI observed that operators specifically asked the model to remove or avoid linguistic indicators that might suggest Russian authorship.
This included producing English-language comments while concealing stylistic or grammatical clues associated with Russian speakers.
This is significant because one of the traditional methods for investigating coordinated influence operations is linguistic attribution.
AI can make that process harder by helping operators:
- Rewrite text naturally
- Translate content
- Remove language-specific errors
- Change tone and vocabulary
- Produce content for multiple platforms
The result is not necessarily better propaganda, but more convincing linguistic camouflage.
The “Sovereignty Index”
A distinctive feature of the operation was a proprietary country-ranking system known as the Sovereignty Index, also referred to as the Burke Index.
The system claimed to evaluate seven categories of sovereignty, including:
- Political
- Economic
- Technological
- Information
- Cultural
- Cognitive
- Military
According to the IBI website, these individual scores were combined into a cumulative national sovereignty ranking.
The published rankings placed the United States first, followed closely by China, with Russia also receiving a relatively high score.
OpenAI assessed that the index was used to present Russia in a favorable light while criticizing Western governments and institutions.
Operation Reached Relatively Small Audiences
OpenAI emphasized that the campaign appears to have reached relatively small audiences.
The largest audiences were associated with Telegram channels, some of which reportedly had approximately 10,000 to 20,000 followers.
This means the operation should not automatically be described as having achieved large-scale political influence.
However, OpenAI argues that its significance lies more in the infrastructure being built than in immediate audience size.
“American Observer” Telegram Activity
The influence campaign was not limited to promoting the International Burke Institute.
OpenAI found that one ChatGPT user generated a profile image for a Telegram channel called American Observer.
The same user repeatedly asked ChatGPT to produce Russian-language summaries of activity on that channel.
This suggests the AI tool was being integrated into a broader workflow rather than used only for individual posts.
AI Was a Supporting Tool, Not the Entire Operation
One of the most important aspects of OpenAI's report is that ChatGPT was not responsible for building or operating the entire campaign.
Instead, AI was used as a supporting layer inside a larger information operation.
The broader infrastructure included:
- Websites
- Purported experts
- Republished academic material
- Social media channels
- Telegram communities
- Custom political indexes
AI mainly helped operators create content that supported and promoted these assets.
This is an important distinction because influence operations do not necessarily need fully autonomous AI agents.
Even relatively simple AI-assisted content creation can improve operational scale and consistency.
Analysis and context
This case demonstrates how generative AI can influence the economics of information operations without necessarily making them more successful.
The most important lesson is not that ChatGPT created a massive propaganda campaign.
It did not.
The more significant development is that AI reduced the amount of human effort needed to maintain a credible-looking influence infrastructure.
AI Helps Manufacture Authority
Influence operations often need more than viral social media posts.
To appear legitimate, they may create:
- Think tanks
- News websites
- Expert profiles
- Research reports
- Rankings
- Academic-style content
These assets create what can be described as manufactured authority.
A social media post linking to an unknown account may have little influence.
The same post linking to a professional-looking “international research institute” can appear more credible.
OpenAI's investigation shows how AI-generated promotional content can support this wider credibility-building process.
AI Can Reduce Linguistic Attribution Signals
Another important lesson is that language is becoming less useful as an attribution signal.
Previously, investigators could sometimes identify likely geographic origins through:
- Grammar mistakes
- Unusual word choices
- Direct translations
- Cultural expressions
Modern generative AI can remove many of those indicators.
A Russian-speaking operator can generate polished English content without being fluent in English.
This does not eliminate attribution entirely, but it reduces the value of purely linguistic indicators.
Defenders increasingly need to combine language analysis with:
- Infrastructure analysis
- Account behavior
- Posting patterns
- Payment information
- Network indicators
- Cross-platform relationships
Reach and Capability Are Different
It is also important not to exaggerate the operation's success.
OpenAI explicitly describes its audience as relatively small.
That means the campaign demonstrated capability and infrastructure, but not necessarily large-scale impact.
This distinction is critical in threat intelligence.
A sophisticated campaign does not automatically mean an effective campaign.
Influence Operations Are Becoming Modular
The campaign also illustrates a broader evolution toward modular influence operations.
Different components can be separated:
Research-looking website
→ establishes credibility
Fake or misattributed experts
→ creates perceived authority
Custom rankings and reports
→ generate shareable narratives
AI-generated social media posts
→ distribute those narratives
Telegram and social platforms
→ provide audiences
This modular architecture makes operations easier to scale and adapt.
AI can accelerate individual components without needing to control the full system.
Detection Can Benefit From AI Too
There is an important defensive side to this story.
The same AI systems that attackers attempt to exploit can also provide detection signals.
OpenAI says its investigation began with suspicious AI-generated social media activity and eventually exposed the much broader operation.
This demonstrates that the use of AI can itself leave traces that help defenders identify coordinated campaigns.
The Broader Security Lesson
The main lesson is that generative AI does not need to produce highly persuasive propaganda to matter.
Its value to influence operators may be much more practical:
- Faster content production
- Multilingual rewriting
- Consistent messaging
- Persona maintenance
- Reduced linguistic fingerprints
- Lower operational costs
The defensive challenge is therefore moving beyond simply detecting “AI-generated text.”
Security teams, platforms, researchers, and governments increasingly need to identify the networks, infrastructure, coordination patterns, and deceptive identities behind the content.
The Russian-linked campaign disrupted by OpenAI shows that AI is becoming a component of influence operations rather than necessarily replacing traditional influence techniques.
Source: OpenAI, Disrupting a new covert influence campaign from Russia, August 25, 2026; additional reporting by The Hacker News, August 26, 2026.