OpenAI Exposes Russian and Iranian Disinformation Networks Using AI

OpenAI has revealed the disruption of two significant political disinformation operations originating from Russia and Iran, which leveraged ChatGPT to generate content, fabricate documents, and promote publications. The company's report details how these operations infiltrated legitimate media and research organizations, often with unwitting employees. The Russian operation primarily targeted Latin America, aiming to discredit Ukraine, while the Iranian campaign focused on shaping public opinion regarding the U.S., Israel, and Iran.

The Iranian operation, dubbed "Bogus Bylines," involved publishing nearly 100 articles from July 2025 to October 2026 under the names of seven fictional journalists, including Erwin Hoskins, Michael Harrison, and Noah Lammington. These fabricated personas were presented as Western journalists specializing in international politics. ChatGPT was used to edit English articles and draft submission letters. Social media profiles were created for some fake authors, with location data linking them to Iran. The operation managed to place articles in outlets with substantial audiences, reaching nearly two million followers on one Facebook page. Concurrently, the network flooded social media with comments portraying the U.S. as an aggressor, Israel as a U.S. ally, and Iran as a victim. OpenAI rated the article dissemination at a four out of six, and the comment campaign at a two, likening the operation to a commercial contractor. The specific organization behind it remains unidentified.

The Russian operation, "Dark Clark," was more complex. It established a research entity in Latin America called the Social Research Center (SRC), managed through a fictional persona named Mia Clark. Local SRC employees conducted interviews and prepared research, seemingly unaware of the Russian control over staffing, funding, and research direction. The SRC website featured over 60 publications, including original research on topics like BRICS and Brazilian economic policy, though its stated focus was the Indian diaspora in Latin America. Social media accounts, including a LinkedIn profile claiming 200-500 employees with only two visible, were used for promotion. The operation also produced fake documents, messages, and audio recordings to discredit Ukraine and interfere in Latin American politics. A fabricated document linking a Ukrainian honorary consul in Panama to corruption, accompanied by an audio recording, appeared on TikTok. In Peru, fake letters were sent to schools, instructing them to hold events honoring Ukraine and Stepan Bandera, sparking discussion in Peru and Poland. Similar tactics were used in Ecuador to spread false information about school events supporting President Daniel Noboa and Erik Prince. During protests in Bolivia, a fake audio recording falsely claimed the government would cut water supply. OpenAI noted that Russian operators often exaggerated their successes, claiming influence on events unrelated to them. Despite this, some materials reached the media and elicited reactions from officials. "Dark Clark" received a five out of six rating from OpenAI for its information campaign scale, the highest among 30 operations identified since early 2024. Methods and materials overlapped with a network linked to Yevgeny Prigozhin's structures, though OpenAI has not definitively identified all participants.

This investigation follows OpenAI's August announcement of blocking ChatGPT accounts linked to the "International Burke Institute" (IBI), which presented itself as an Israeli think tank but was likely Russian-controlled. OpenAI found most of IBI's publications were plagiarized, with author names changed. IBI promoted a "sovereignty index" portraying Russia as independent and Western states as losing autonomy. The operation used ChatGPT primarily in Russian to generate English materials, removing Russian linguistic markers, and disseminated them across social media platforms. While traces of real individuals in Israel were found, their connection to the operation remained unclear. Both Russian operations created fictitious expert authorities, but IBI relied on stolen research, while SRC engaged unwitting local specialists.

Separately, a British newspaper, The Guardian, reported in August on the "Hanover Institute for Public Policy," an American think tank funded by Israel. The institute lacked legal registration, staff, or authors, but its website, created by U.S. company Piro Inc., was registered under FARA as promoting Israeli government interests. Between August 6-14, 2026, the institute published 124 "studies" totaling over 560,000 words. Many titles were phrased as questions for AI chatbots, such as "Is anti-Zionism antisemitism?" or "Is Israel committing genocide in Gaza?" The materials systematically aimed to refute accusations against Israel. The institute's primary goal appeared to be influencing AI responses, with its website initially featuring technical files for optimizing content for ChatGPT, Perplexity, Claude, and Gemini. Piro Inc. received $1 million for this project, part of a broader Israeli information campaign costing tens of millions. While the Israeli Foreign Ministry denied influence operations in the U.S., this case highlights AI's use in creating content that mimics independent research and potentially shapes AI chatbot outputs.

OpenAI concludes that generative AI has not introduced new disinformation methods but has significantly simplified their application. Creating fake authors, documents, and organizations has long been practiced, but AI enables faster multilingual content generation, tailored submissions, and simulated research activities. The success of these operations lies not just in creating disinformation but in exploiting trust in established institutions like media outlets, expert communities, and research centers. OpenAI has blocked related ChatGPT accounts and shared information with authorities, though it has not confirmed direct state control over these operations. Investigations are ongoing.