OpenAI Bans Accounts Behind Two Influence Operations That Used ChatGPT to Fake Journalists and a Think Tank

OpenAI said on Thursday that it had banned a cluster of ChatGPT accounts behind two covert influence operations, one originating in Russia and one in Iran, that used its models to fabricate journalist identities, run a fake think tank and place political content in real publications. The findings come from the latest entry in the company's reporting line on influence operations.

The Iranian operation, which OpenAI dubbed Bogus Bylines, built seven fictitious journalist personas presented as freelance writers from Western countries. The personas pitched long-form articles about tensions between the United States and Iran to online outlets around the world. OpenAI identified almost 100 articles published or syndicated under those bylines across roughly a dozen small and medium outlets focused on international affairs and the Middle East, with the earliest identified article from July 2025 and the most recent from October 2026. One persona, using the name Ervin B. Hoskins, maintained accounts on X and Instagram; transparency data shown by OpenAI indicated the X account connected through a West Asia Android app and the Instagram account was based in Iran. OpenAI rated the article placement at Category 4 on the Brookings Breakout Scale, and the operation's social media commenting, which drew little engagement, at Category 2.

The Russian operation, named Dark Clark, was more elaborate. It ran through an organization called the Social Research Center, which presented itself publicly as an independent research platform covering the Indian diaspora in Latin America. According to OpenAI, Russian operators controlled the group through a fabricated identity and recruited local employees who appear to have been unaware they were working for a Russian operation. The center produced fabricated reports, forged emails and fake audio recordings aimed at undermining the reputation of Ukraine in the region and influencing political debates in countries including Argentina and Bolivia. Some of its claims prompted fact checks or official denials. OpenAI assessed Dark Clark at Category 5, the highest level it has assigned to any operation it has disrupted since it began publishing these reports.

What stands out in the report is how conventional the tradecraft remains. OpenAI wrote that the operations resemble influence campaigns from before generative AI, using AI mainly to draft reports, polish articles to match the editorial style of specific outlets, and prepare pitch emails, rather than to automate deception end to end. The company also cautioned that the operators' own claims about their impact should not be taken at face value, since internal reports from influence campaigns tend to exaggerate results.

OpenAI said it banned the ChatGPT accounts associated with both operations and shared information with relevant authorities and industry partners. The company has published similar reports for two and a half years, and this is the first time it has assigned a Category 5 rating to a disrupted operation.

For newsrooms and platform trust teams, the practical lesson is about contributor provenance rather than prose style. A plausible freelance byline with social accounts, a clean pitch email and fluent copy can still be part of a coordinated campaign, and the identifiers that give it away are more likely to be account metadata and cross-platform patterns than anything in the text itself.