Article Details

OpenAI kills Iranian accounts using ChatGPT to write US election disinfo. 12 on X and one on Instagram caught in the crackdown. OpenAI has banned ChatGPT accounts linked to an Iranian crew suspected of spreading fake news on social media sites about the upcoming US presidential campaign. The accounts, we're told, used ChatGPT to generate long-form articles to share on social media platforms, as well as short social comments about both candidates running for America's top office: Democrat Kamala Harris and Republican Donald Trump. Despite being posted on several websites, however, the influence operation "does not appear to have achieved meaningful audience engagement," according to a Friday alert. OpenAI attributed the phony posts to Storm-2035, a Tehran-backed group that Microsoft also sounded the alarm about last week as it and other Iranian groups have continued to meddle in elections — some veering toward attempts at inciting violence. Later last week, Google's threat hunters published intel on Iranian cyber influence activity following a recent uptick in attacks that led to data being leaked from the Trump re-election campaign. In total, the AI org says it identified 12 accounts on X and one on Instagram involved in this covert influence operation. In addition to the US presidential election, the made-up news and comments also covered the conflict in Gaza and Israel's participation in the Olympics, as well as other politically charged topics.  "The first workstream produced articles on US politics and global events, published on five websites that posed as both progressive and conservative news outlets," OpenAI said on Friday. "The second workstream created short comments in English and Spanish, which were posted on social media." The five domains are: Storm-2035, and at least one of its fake-news sites EvenPolitics, has been active online since the US midterm elections in 2022, publishing about 10 "articles" a week. While these are written in English, the group also has other influence-operation websites in  Arabic, English, French and Spanish. Based on Brookings' Breakout Scale, which rates the effectiveness of these types of covert operations on a scale from 1 (lowest) to 6 (highest), Storm-2035 only garnered a 2, meaning it was active on multiple platforms but showed no real evidence of being picked up or widely spread by real people. "The majority of social media posts that we identified received few or no likes, shares, or comments," OpenAI noted. "We similarly did not find indications of the web articles being shared across social media." Earlier this year, OpenAI shut down five accounts it said were being used by government-backed groups in China, Iran, Russia and North Korea to generate phishing emails and malicious software scripts.