OpenAI says it has dismantled two covert influence operations that used its generative AI tools to fabricate journalist personas and front think tanks aimed at shaping coverage in Western and Latin American media.[2][8][11] In a report released Thursday, the company said it terminated clusters of accounts linked to Russia and Iran and rated the campaigns 5 and 4 on its internal six-point impact scale, a substantially higher score than the low-impact operations it usually observes.[2][8]
The Russian-origin network, dubbed “Dark Clark,” revolved around a fabricated Social Research Center (SRC), a supposed Latin American think tank fronted by a fake persona named Mia Clark that operators built and managed with assistance from ChatGPT.[1][11][13] According to OpenAI’s investigation, summarized by outlets including Unite.AI and TokenPost, the cluster of accounts used the company’s models to draft internal reports, set wage scales, and make hiring decisions, indicating that Russian operators were likely directing the SRC’s activities rather than merely infiltrating an existing organization.[1][13] OpenAI assessed that employees of the SRC working in Latin America appeared unaware they were effectively part of a Russian influence effort, and that much of the think tank’s output was original content produced by these co‑opted staff members.[11][13]
Dark Clark’s messaging focused on undermining support for Ukraine and reshaping perceptions of the war across Latin America, while also injecting itself into sensitive domestic debates in Argentina and Bolivia.[1][11] Fact-checkers in the region, including Lupa in Brazil, documented how the network amplified a TikTok video falsely alleging that Ecuador’s government was using shell companies to recruit citizens to fight for Ukraine, and spread fabricated audio purporting to show the Ukrainian consul in Ecuador insulting Ecuadorian nationals, prompting official denials.[1][11] During anti-government protests in Bolivia, the same operation circulated another fake audio clip in which a supposed employee of the state water company claimed authorities were about to declare a state of emergency and cut off water to La Paz, again forcing the Bolivian government to issue a public rebuttal.[1][11] OpenAI described Dark Clark as the most complex front organization it has disrupted in more than two years of monitoring influence campaigns.[11][13]
The second campaign traced to Iran used similar AI tooling to run what OpenAI characterizes as a for‑hire media influence operation built around at least seven fictitious freelance journalists.[2][8] Operators leveraged ChatGPT to pitch and produce articles about U.S.-Iran tensions and broader Middle East geopolitics, successfully placing and syndicating almost 100 stories across roughly a dozen small and mid-sized outlets worldwide focused on international affairs.[2][8] Coverage in Coingape and other publications notes that the articles often attracted little reader engagement or social media amplification, underscoring that even relatively successful AI-assisted covert operations still face challenges in organically reaching large audiences.[2][8] OpenAI stopped short of attributing the network directly to the Iranian state, saying the pattern of activity more closely resembled a commercial entity offering influence services.[2][7]
These takedowns build on OpenAI’s earlier disclosures from 2024, when the company detailed how actors linked to Russia, China, Iran and Israel had begun folding generative models into existing disinformation pipelines.[3][5][6] In that report and subsequent coverage by outlets such as The Record, OpenAI highlighted Russian campaigns like “Bad Grammar” and “Doppelganger,” as well as Iran’s International Union of Virtual Media, which used AI to generate comments, articles and headlines but achieved limited real-world impact.[6][9][10] While the company has consistently argued that AI has so far provided only marginal amplification for most influence operations, the unusually high impact scores assigned to Dark Clark and the Iranian journalist network suggest that more sophisticated use of front organizations and believable personas is beginning to help some campaigns break into mainstream discourse.[2][5][12]
For defenders in government, media and civil society, the cases underscore how generative AI can accelerate traditional influence tradecraft by lowering the cost of manufacturing plausible experts, think tanks and localized narratives at scale.[3][6][10] OpenAI’s findings, echoed across multiple independent analyses, point to the need for newsrooms and platforms to scrutinize sudden appearances of new “research centers” and prolific freelance contributors, to verify identities and funding sources, and to respond quickly with fact checks when dubious claims gain traction.[1][2][11] At the same time, the company notes that most AI-assisted campaigns it tracks still languish at the bottom of its impact scale, suggesting that transparency from platforms, rapid account takedowns, and resilient local media ecosystems can meaningfully blunt the effects of even increasingly sophisticated AI-enabled propaganda.[2][6]
References
- OpenAI Bans Two Covert Influence Operations Using False …
- OpenAI Reports Influence Operations of Russian, Iranian …
- OpenAI Says Russia and China Used Its A.I. in Covert …
- The era of ChatGPT-powered propaganda is upon us
- OpenAI models used in nation-state influence campaigns …
- OpenAI Disrupts Iranian Covert Influence Campaign
- OpenAI Bans Russia and Iran-Linked Accounts Over Secret Influence Operations
- OpenAI disrupts disinformation operations tied to China, Russia
- OpenAI accuses Russia, China, Iran, and Israel of misusing its …
- OpenAI: Русия и Иран използват ChatGPT, за да разпространяват лъжлива информация
- In a first, OpenAI removes influence operations tied to …
- OpenAI Blocks Two AI-Assisted Influence Operations Using Fake Identities | TokenPost
