OpenAI has banned clusters of ChatGPT accounts linked to Russian and Iranian influence operations that used artificial intelligence to produce political content, create deceptive online identities and spread misleading narratives through websites and social media. The company disclosed the operations on October 8, 2026, describing how the groups combined AI-generated material with traditional influence techniques to make their campaigns appear legitimate. The Russian operation targeted audiences across Latin America, while the Iranian campaign used fabricated journalist identities to place articles in established online publications. OpenAI said it had reported relevant findings to authorities and disrupted the accounts involved. OpenAI’s investigation details the methods used in both campaigns.
The disclosure highlights a growing challenge for the technology industry: generative AI can help people produce content quickly, but the same capabilities can be exploited to manufacture credibility and amplify political messaging. OpenAI found that the Iranian operation had placed nearly 100 articles under fabricated bylines across more than a dozen publications, while the Russian-linked operation appeared to use a purported research organisation to promote geopolitical narratives and influence political discussions. The findings do not mean that every article generated with AI is misinformation or that every publication involved knowingly participated in a campaign. Instead, they show how AI-generated content can enter legitimate media channels when deceptive identities and misleading claims evade editorial checks.
OpenAI Identifies Two Covert Influence Operations
OpenAI described the campaigns as “false front” operations because the people and organisations presenting the messages appeared independent of the actors behind them.
Rather than relying exclusively on fake social media accounts, the operators used ChatGPT alongside websites, fabricated identities, editorial pitches and other conventional techniques. Their objective was to place geopolitical messaging in front of audiences through channels that might otherwise appear credible.
The two operations differed in their methods and apparent targets.
| Operation | Origin identified by OpenAI | Main approach | Primary focus |
|---|---|---|---|
| Dark Clark | Russia | Purported research organisation, reports and online influence material | Latin America, Ukraine and local political issues |
| Bogus Bylines | Iran | Fake journalist identities and articles pitched to online publications | US-Iran relations and the wider conflict |
OpenAI said it banned the associated account clusters after investigating their activities. The company also shared information with relevant authorities.
These findings are based on OpenAI’s own threat-intelligence investigation. Attribution of covert campaigns can be difficult, and the company described the operations as originating in Russia and Iran rather than establishing that every individual involved acted under direct government instructions.
Russian Operation Used a Purported Latin American Think Tank
The Russian-linked campaign, which OpenAI called Operation Dark Clark, used ChatGPT to support activities targeting audiences in Latin America.
According to OpenAI, the operators generated content and internal reports connected to a purported research platform called the Social Research Center (SRC). The organisation presented itself as a research initiative, but OpenAI said it was part of a broader influence operation.
Much of the activity appeared designed to undermine Ukraine’s reputation in the region. Some material also appeared aimed at influencing local political outcomes, particularly in Argentina and Bolivia.
How the operation worked
OpenAI identified three broad uses of ChatGPT in the campaign:
- Internal reporting: Producing reports about the operation’s activities and the work of people associated with it.
- Content creation: Generating material intended to support the campaign’s messaging.
- Research-platform reporting: Drafting reports about the performance of the purported Social Research Center.
The operation also used conventional online influence techniques. Its reliance on a purported research organisation was intended to give political messaging the appearance of independent analysis rather than openly partisan propaganda.
OpenAI said the operation appeared to have involved people in Latin America who were unaware of the Russian operators behind the project. This is an important distinction: individuals associated with a purported organisation may not necessarily know that they are contributing to a covert influence campaign.
The company described Dark Clark as its first disrupted influence operation rated Category 5 on its six-level Breakout Scale. That rating reflects its assessment of the operation’s characteristics and apparent impact, rather than a universal measure of all Russian disinformation activity.
Iranian Campaign Used Seven Fake Journalist Identities
The second operation, named Bogus Bylines, involved accounts that OpenAI traced to Iran. The operators used ChatGPT to draft and refine long-form articles, editorial pitches, social media comments and internal reports.
The campaign relied on seven fabricated journalist personas. These identities were presented as writers from Western countries, helping the operators pitch material to publications that might not otherwise have considered content from the actual people behind the operation.
OpenAI identified nearly 100 articles published or syndicated under the campaign’s bylines across more than a dozen online outlets worldwide. Some of the articles focused on the US-Iran conflict and presented narratives critical of US actions.
The operators generated material in Persian and English, and used virtual private networks (VPNs) to obscure their location when accessing ChatGPT.
Why publishing through real outlets matters
Publishing an article on an established website can give a message a degree of credibility that a newly created propaganda website may struggle to achieve.
Once an article appears on a legitimate publication, it can be shared through social media, cited by other websites or encountered by readers who have no reason to suspect that its author identity is fabricated.
OpenAI said at least some articles were subsequently promoted through the social media accounts of the outlets that published them. This gave the campaign a potential route to audiences beyond those who would have encountered the original content.
The company assessed the Iranian operation as Category 4 on its six-level Breakout Scale.
The findings do not establish that every publication involved knowingly supported the operation. In many cases, the central deception concerned the identities and affiliations of the people pitching the material.
How AI Helped the Campaigns
Neither operation relied on AI alone. ChatGPT was used alongside conventional websites, social media accounts, editorial outreach and other methods of distributing information.
Nevertheless, generative AI can reduce the effort required to produce large volumes of text, adapt messages for different audiences and revise material to fit particular editorial styles.
| AI-supported activity | Potential advantage for influence operators |
|---|---|
| Drafting articles | Producing long-form material more quickly |
| Editing and translation | Adapting messages for different languages and audiences |
| Editorial pitches | Preparing customised proposals for publications |
| Social media comments | Creating multiple variations of similar messages |
| Internal reporting | Tracking activity and summarising campaign results |
| Persona development | Supporting content presented under fabricated identities |
These capabilities are not inherently malicious. Translation, editing, article drafting and social media management are legitimate uses of AI.
The problem arises when the tools are used to support deception, impersonation or covert manipulation. The campaigns illustrate how legitimate publishing channels can be exploited when content production is combined with false identities and misleading claims about who is speaking.
OpenAI Says Media Placement Reached Wider Audiences
One of the more significant findings was OpenAI’s assessment that campaigns placing misleading narratives in established media outlets appeared to reach wider audiences than operations focused primarily on fake social media accounts.
This distinction matters because influence operations are not successful simply because they generate large quantities of content. Their impact also depends on whether the material reaches people, appears credible and is shared or repeated by others.
A fabricated article published on a recognised website may be more influential than dozens of posts on accounts with little visibility. It can also be separated from the original campaign and circulated without the context needed to identify its source.
OpenAI’s findings demonstrate potential reach, not proof that every reader believed the material or changed their political views because of it. Measuring actual persuasion is considerably harder than identifying content and tracing its distribution.
The company did not claim that AI alone made the campaigns successful. Instead, its investigation showed how AI assistance could complement existing techniques for creating and distributing deceptive political messaging.
What OpenAI Is Doing to Counter AI Misuse
OpenAI said it banned the ChatGPT account clusters linked to both operations and shared relevant information with authorities.
The company has also been publishing threat-intelligence reports describing how malicious actors use its tools. These reports are intended to help policymakers, researchers, technology companies and the wider public understand emerging methods of AI-enabled abuse.
However, account bans address only one part of the problem. Operators may attempt to create new accounts, use different AI services or switch to other content-generation methods.
Detecting such campaigns also requires examining behaviour beyond the AI platform itself. Researchers may need to investigate websites, social media profiles, publishing histories, account relationships and other evidence to identify coordinated activity.
For media organisations, the incidents reinforce the importance of verifying author identities, checking sources, confirming claims and scrutinising unsolicited pitches. A well-written article is not necessarily an accurate article, and fluent language cannot establish the credibility of its author.
The Bigger Picture
OpenAI’s findings show how generative AI can be incorporated into sophisticated influence operations that rely on more than fake social media posts. The Russian-linked campaign used a purported research organisation, while the Iranian operation relied on fabricated journalist identities and real publishing outlets. Both demonstrate how misleading political narratives can gain credibility when the identity and intentions of their creators are concealed.
The challenge for the technology and media industries is to distinguish legitimate AI-assisted communication from organised deception. Account bans, threat intelligence and stronger editorial verification can help, but no single intervention will eliminate the problem. Effective responses will require cooperation among AI developers, publishers, researchers and authorities, alongside careful protection of legitimate political expression.
Looking Ahead
OpenAI is likely to continue investigating coordinated misuse of its models as influence operators adapt their methods. Future assessments will need to examine whether account enforcement disrupts campaigns for meaningful periods, whether the same actors move to other tools and how widely their content spreads after publication. Media organisations will also face pressure to strengthen verification practices without treating all AI-assisted writing as suspicious.
Over the longer term, the incidents underline the importance of digital provenance, transparent authorship and stronger verification of political content. AI can make it easier to create convincing articles and adapt messages across languages, but it cannot establish whether the underlying claims are true. For readers, publishers and technology companies alike, the essential task is to evaluate both the content and the credibility of the people or organisations presenting it.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
