TL;DR

OpenAI said on Thursday 8 October that it has banned ChatGPT accounts linked to two covert influence operations, a Russian one and an Iranian one, which relied on fake organisations and personas to slip propaganda into real media. It rates the Russian campaign, which it calls “Dark Clark”, Category 5 on the six-point IO Breakout Scale, the first it has disrupted at that level.

Dark Clark: a think tank with unwitting staff

The Russian operators ran a so-called research platform in Latin America, the Social Research Center, through a fake persona named “Mia Clark”. Their own reports discussed hiring, firing and pay, and OpenAI says the local employees appear not to have known who they were working for. The centre published well over 60 articles, mostly original work.

The operators used ChatGPT chiefly to draft progress reports for an unnamed boss, rather than to create campaign material. Those reports describe efforts to undermine Ukraine and leaders in Argentina, Bolivia and Ecuador. In one case, the operators say a fake email posing as Lima’s education authority got Peruvian schools to hold events referencing Stepan Bandera; OpenAI found matching stories in the Peruvian and Polish press. Fact checks and official denials in several countries suggest the fakes reached real audiences, though OpenAI warns the operators also claimed credit for things they did not do.

Bogus Bylines: fake journalists, real outlets

The Iranian network pitched long-form articles on the US-Iran conflict under seven invented Western journalist personas, using ChatGPT to fit each draft to an outlet’s submission rules and write the pitch email. OpenAI found almost 100 articles under those bylines in around a dozen small and medium outlets, one of which has close to 2 million Facebook followers. It rates that strand Category 4.

The same operators generated batches of hostile comments, including more than two dozen sets of Persian-language replies aimed at posts by Iran International, the UK-based satellite channel critical of Tehran. Those comments drew little engagement. OpenAI says the activity looked like a commercial, for-hire campaign, but could not identify who ran it.

Looking forward

OpenAI notes that both operations echo pre-AI fronts such as the fake journalist “Alice Donovan” and the outlet “PeaceData”, both of which stopped once exposed. AI made the work faster and more fluent, not new. In our view, the lesson for UK editors and publishers who accept contributed articles is that a polished pitch from an unfamiliar freelancer now proves little, and checking a writer’s identity matters more than checking their prose. On the same day, Anthropic gathered its own rules against deceptive campaigns into a single policy section.