Loading page…
The company’s investigation points to fake official emails and apparently unwitting research staff—not just AI-generated propaganda—as routes to public attention.
Listen to this story
OpenAI rated the Dark Clark operation Category 5 on its six-level impact scale, the first campaign it says it has disrupted to reach that level. ChatGPT was used mainly to draft internal reports, while alleged fake school directives in Peru and Ecuador generated real-world responses and matching media coverage; a front called the Social Research Center also appears to have relied on local staff unaware of its Russian operators. OpenAI said public evidence corroborates parts of the operators’ accounts, not every claimed outcome.
OpenAI said Dark Clark used VPNs to access ChatGPT, which it blocks from Russia.
In Peru, fake education-authority emails urged schools to hold events referencing Stepan Bandera; some schools allegedly responded with photos.
In Ecuador, operators claimed they directed schools to pledge allegiance to President Daniel Noboa and Erik Prince; the education minister issued a detailed rebuttal.
Schools and news outlets helped spread a Russia-origin influence campaign, OpenAI said in an October 8, 2026 disclosure. The company banned accounts tied to the operation, nicknamed Dark Clark. Its investigation described fake official instructions and apparently unwitting research staff helping spread messages that sought to undermine Ukraine and influence Latin American politics.
OpenAI called it the most far-reaching influence operation it had disrupted, NBC News reported. The company assigned it Category 5 on a six-level scale of influence-operation impact—the first campaign it has disrupted to reach that category.
The operators’ main use of ChatGPT was drafting and updating internal reports to an unknown superior. Those reports covered anti-Ukraine campaigns, interference in domestic politics, and management of a self-described research platform. They also described other people’s activities that the operators could plausibly claim as their own.
The accounts used virtual private networks, or VPNs, to access the service. OpenAI blocks model access from Russia.
In Peru, the operators claimed they impersonated Lima’s Regional Directorate of Education through a fake email address in May 2026. They instructed schools to hold Ukraine-themed events mentioning controversial nationalist Stepan Bandera. Some schools allegedly replied with confirmation and photographs, which the operators said they used to plant stories portraying Ukraine as exporting ultra-nationalist ideology.
OpenAI found matching coverage in Peru, Poland and an English-language publication in Hungary. Some articles carried a Polish European Parliament member’s call to bar people displaying “anti-Polishness” from Poland. The company said the fake fed into historical tensions between Ukrainians and Poles.
The operators claimed a similar email deception in Ecuador in June, directing schools to pledge allegiance to President Daniel Noboa and former Blackwater head Erik Prince. OpenAI identified closely matching local coverage and a detailed rebuttal from Ecuador’s education minister. Those public traces support parts of the operators’ accounts, not every claimed result.
Dark Clark also appears to have run the Social Research Center in Latin America through a fake persona named Mia Clark. Internal reports discussed hiring, firing and pay, suggesting the operators controlled the center rather than worked with it from outside. OpenAI said local employees apparently did not know they worked for a Russian group. Those employees produced most of the center’s original content.
OpenAI said it shared information about the case with relevant authorities.
Story updates
Real publications carried articles from fabricated journalists in an Iran-origin influence campaign that used ChatGPT to prepare submissions, OpenAI disclosed on October 8. The Washington Post’s October 9 reporting traced one route into a Florida weekly: a trusted contributor organization whose relationship with the paper substituted for independent checks on the writer.
OpenAI named the operation Bogus Bylines. Its operators used seven fake journalist personas to pitch long-form articles about the US-Iran conflict to small and medium online publications around the world. OpenAI said the operation originated in Iran but could not identify the specific actor behind it.
ChatGPT’s role extended beyond producing copy. According to Digital Trends’ account of OpenAI’s findings, operators supplied English-language drafts and asked the chatbot to check them against a publication’s submission requirements. It suggested edits; once the operators approved those changes, they requested an email pitching the editor.
The fabricated identities also had supporting material. Some maintained social accounts, and OpenAI said ChatGPT wrote one persona’s biography. These were not simply anonymous texts posted online: the operation paired articles with apparent authors and tailored approaches to the publications it wanted to reach.
The Post identified an opinion piece published in September in the Port St. Joe Star, a small Florida weekly, under the name Ervin Hoskins. It warned about the costs of the Iran war. The same byline appeared on articles critical of the war in Daily Kos and Middle East Monitor.
OpenAI’s investigation added evidence about the Hoskins identity. It said the Ervin B. Hoskins persona maintained accounts on X, Facebook, Instagram and Threads. Social-platform transparency information located those accounts in Iran. That evidence supports the company’s geographic attribution; its inability to name a specific actor remains a separate limit.
In its October 8 disclosure, OpenAI said it identified almost 100 articles published or syndicated across roughly a dozen online outlets worldwide. The earliest linked article it identified appeared in July 2025 and the latest in October 2026, putting the identified publishing activity across more than a year.
The Post gave a different outlet total: about 100 AI-generated articles across at least 20 news outlets, attributing those figures to OpenAI. The article totals are similar, but the outlet counts conflict. Neither figure should be treated as a measure of how many readers were persuaded by the campaign.
OpenAI banned the accounts involved and said it shared information with relevant authorities. Some publications later removed articles, and PeaceVoice tightened its AI policy, according to the Post’s reporting summarized by Digital Trends.
openai.comLoading discussion...
Join the conversation
Explain whether the tool’s role changes how you judge the risk.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.