OpenAI says it banned account clusters tied to Russian and Iranian influence operations that used ChatGPT to prepare internal reports, refine articles and create some deceptive material. Its October 8 account describes campaigns built around fake institutions and identities, while leaving the scale of their real-world audience uncertain.

The big change

  • The report presents ChatGPT as one tool in a broader influence workflow, with its clearest role in internal reporting and editing.
  • OpenAI says some fabricated material appeared online and that Iranian operators placed articles under fake bylines in real outlets; the company’s own account is the main source for those findings.
  • The operators’ claimed reach was unreliable. OpenAI says some claims were corroborated, others were not, and audience impact cannot be read from the operators’ own metrics.

The Russian operation, which OpenAI named “Dark Clark,” used ChatGPT mainly to draft and update reports about the campaign, according to the company. OpenAI says the operators also used the model to prepare some fake documents, audio scripts and messages. The operation allegedly worked through a “think tank” called the Social Research Center, fronted by the fabricated persona “Mia Clark,” while people hired in Latin America may not have known who controlled the organization.

OpenAI says its investigators found examples of matching content and reporting by fact-checkers or officials. One example concerned a fabricated video that claimed Ecuador was recruiting young people to fight in Ukraine. An Ecuadorian fact-checking organization, Lupa Media, separately reported that the purported Teleamazonas report did not exist, the video used stolen social-media images, its audio was AI-generated, and the named company had no valid registration. That corroborates the existence and fabrication of a similar item; it does not independently establish who commissioned it or prove every element of OpenAI’s account.

The Iranian operation, called “Bogus Bylines,” used ChatGPT to edit long-form articles and pitches, draft outreach emails, generate batches of social-media comments and polish internal reports, OpenAI says. The company identified almost 100 articles published or syndicated under seven invented journalist identities across roughly a dozen outlets, from July 2025 through October 2026. Those counts are OpenAI’s investigation findings. The independent coverage reviewed for this story reports the company’s findings but does not present a separate audit of all the bylines or publications.

The report also draws a sharp line between producing material and distributing it. OpenAI says the Iranian operation’s social comments were posted close together, but says it found no evidence that its models automated the posting. It describes many comments as attracting little engagement. The operators’ internal reports counted views on the posts they replied to, rather than views on their replies, and compared totals across an unspecified interval. That method cannot isolate whether the campaign caused the views.

OpenAI rated the Russian operation Category 5 and the Iranian article-placement effort Category 4 on the Breakout Scale. The scale, proposed by Brookings in 2020, categorizes influence operations by observable spread across platforms and communities, including whether content crosses into mainstream media or prompts high-profile attention or concrete action. These ratings are OpenAI’s assessments, not independent measurements of how many people believed the messages or changed their behavior.

That distinction matters for readers trying to understand what generative AI changed. The report describes assistance with drafting, editing, translation, scripts and internal summaries. It does not show that ChatGPT independently ran the campaigns, selected targets or controlled their distribution. People and conventional infrastructure supplied the false identities, outlet pitches, social accounts and posting. In one described case, operators used the model to refine a fake letter and translate it, then a matching image and audio appeared online. That sequence illustrates how a text model can support a chain of human-directed production; it does not establish that the model created every part of the final artifact.

The harder problem for platforms and publishers is therefore provenance and coordination, not just detecting machine-written sentences. A polished pitch can be ordinary writing assistance; its risk changes when it is tied to a fictitious byline, hidden sponsor or coordinated placement. Likewise, a real employee can unknowingly produce authentic reporting for an organization whose control they do not understand. The OpenAI report’s account of the Social Research Center makes that possibility especially relevant, though the company’s evidence remains the basis for that characterization.

For newsrooms, the practical checks remain familiar: verify the author’s identity and organizational control, contact sources through independently established channels, inspect claims and documents at their origin, and avoid treating engagement counts as proof of influence. The report supplies examples of how those checks can be defeated or manipulated; it does not provide a complete measurement of exposure or downstream political effects.

OpenAI says it banned the account clusters and shared information with relevant authorities and industry partners. The public report gives readers a detailed view of the company’s investigation, but its categories and campaign attributions should be read as company findings. Independent fact-checking supports some specific examples. The total reach, the operators’ full control structure and any lasting effects remain less certain than the activity OpenAI says it observed.

Sources & further reading