OpenAI bans Russian and Iranian false-front influence operations
OpenAI removes accounts behind two influence campaigns
OpenAI said it banned clusters of ChatGPT accounts connected to two covert influence operations: the Russia-origin campaign it calls Dark Clark and the Iran-origin campaign it calls Bogus Bylines. Both combined AI-assisted work with established deception techniques, including false personas, fabricated material and attempts to place messaging before external audiences.
OpenAI assessed Dark Clark as Category 5 on its six-level IO Breakout Scale, its first reported disruption at that level. The Iranian operation’s article-placement activity was assessed as Category 4, while its social-media commenting activity was assessed as Category 2. The scale measures an operation’s apparent ability to break out beyond its own assets and reach audiences.
Russian operation used a Latin American research front
OpenAI said the Russia-linked operators targeted countries across Latin America and appeared to seek to undermine Ukraine’s reputation, as well as influence domestic political outcomes in Argentina and Bolivia. The group used VPNs to access ChatGPT, because OpenAI does not permit access to its models from Russia.
Most of the Russian operators’ use of ChatGPT involved drafting and updating internal reports for an unknown superior. They also generated campaign material and reports on the Social Research Center, a self-described Latin American research platform that OpenAI said appeared to be controlled through a fake persona named Mia Clark.
The available evidence indicated that staff working for the Social Research Center in Latin America did not know they were working for a Russian operation. OpenAI identified more than 60 articles on the centre’s website, most of which appeared to be original work produced by co-opted staff. Internal reports referred to hiring, firing, pay scales and research projects, indicating control rather than a loose partnership.
OpenAI also linked the campaign’s claimed activity to fabricated emails, audio and documents designed to trigger local reaction. Some matching claims were later fact-checked or denied by public bodies. The company cautioned that the operators also tried to take credit for events they had not caused, so their internal impact claims could not be accepted at face value.
Iranian campaign relied on deceptive bylines
The Iran-origin operation used seven purported journalist identities to pitch and publish long-form English-language articles, largely focused on the US-Iran conflict. OpenAI identified almost 100 articles published or syndicated under those bylines across roughly a dozen small and medium online outlets worldwide.
Its operators used ChatGPT to refine articles against outlets’ submission criteria and draft pitch emails to editors. They also created batches of comments in English and Persian. OpenAI found little evidence that the comments attracted meaningful engagement, and said the campaign’s internal reporting used a misleading metric based on views of the original posts rather than views of its replies.
For publishers, researchers and communications teams, the cases underline the need to verify contributor identities, examine provenance for externally submitted material and treat polished language as insufficient evidence of authenticity before content is published or amplified.