OpenAI has banned a cluster of ChatGPT accounts connected to an Iranian influence operation that produced content about the U.S. presidential election. The company said the activity included AI-generated articles and social media posts, but there was no sign that the operation reached a large audience.
The case shows how generative AI tools are becoming part of the same online influence playbook that has previously relied on social platforms. The apparent goal was not simply to push one side of a debate, but to create political friction across sensitive issues.
What OpenAI Says It Found
OpenAI said the banned accounts were tied to a group Microsoft Threat Intelligence identifies as Storm-2035. According to Microsoft, Storm-2035 is an Iranian network that has been part of a broader effort to influence U.S. elections operating since 2020.
The operation used ChatGPT to create long-form articles and rewrite political comments for social media. OpenAI identified five website fronts connected to Storm-2035. These sites presented themselves as both progressive and conservative news outlets, using domain names that were meant to look credible, including “evenpolitics.com.”
That structure matters because influence operations often try to appear native to the communities they are targeting. A site that looks like a partisan news outlet can give manufactured content a place to live before it is pushed into social media conversations.
OpenAI also found a social media layer. The company identified a dozen X accounts and one Instagram account controlled by the operation. ChatGPT was used to rewrite political comments that were then posted on those platforms.
The Content Targeted Divisive Issues
Microsoft said Storm-2035 used multiple sites imitating news outlets while “actively engaging US voter groups on opposing ends of the political spectrum with polarizing messaging on issues such as the US presidential candidates, LGBTQ rights, and the Israel-Hamas conflict.”
That description points to a familiar pattern. The operation did not need to persuade every reader of a single viewpoint. It could instead try to sharpen conflict by placing charged content in front of different audiences.
OpenAI said the websites included long-form articles drafted with ChatGPT. One article alleged that “X censors Trump’s tweets,” even though the source article notes that Elon Musk’s platform has not done that and that Musk is encouraging former president Donald Trump to engage more on X.
On social media, one post falsely claimed that Kamala Harris attributes “increased immigration costs” to climate change, followed by “#DumpKamala.” OpenAI described this as part of the content rewritten with ChatGPT and then posted through accounts controlled by the operation.
Why Generative AI Changes The Cost Of Influence
The important shift is not that influence operations exist. The source article draws a connection to earlier attempts by state actors to use platforms such as Facebook and Twitter during previous election cycles. What changes with tools like ChatGPT is how quickly text can be generated, revised, and adapted for different audiences.
AI-generated articles and posts can help an operation create the appearance of activity across websites and social channels. A network can build several fronts, write in different political voices, and keep posting without needing the same level of manual writing effort.
Based on the source, OpenAI’s response has been to remove accounts when it identifies them. The company previously disrupted five campaigns in May that used ChatGPT to manipulate public opinion. This latest ban adds another example of OpenAI acting against accounts linked to state-affiliated misuse.
That approach can reduce the reach of a known cluster, but it also suggests an ongoing enforcement problem. If these efforts are quick and cheap to set up, platforms may continue finding new groups after they have already produced websites, articles, and social posts.
The Reach Appeared Limited
OpenAI said it did not find evidence that Storm-2035’s articles were widely shared. It also said a majority of the operation’s social media posts received few to no likes, shares, or comments.
That limited engagement is an important part of the story. The operation produced content and controlled accounts, but the available evidence described in the source does not show a successful mass audience. In influence work, volume alone does not prove impact.
Still, limited reach does not make the activity irrelevant. The low cost of producing AI-generated political content means many such attempts can be launched even if most fail to attract attention. A small number of posts can disappear into the feed, while the broader pattern still forces platforms to spend time investigating and removing coordinated activity.
For readers, the practical lesson is simple: election-related content can now come from networks that imitate news outlets, post from multiple accounts, and use AI tools to reshape political messages. The strongest signal in this case is not one viral post, but the combination of fake media fronts, social accounts, and automated writing support.
What To Watch Next
The source article suggests that more notices like this are likely as the election approaches and online partisan conflict intensifies. OpenAI has already acted against multiple campaigns, and the Storm-2035 case shows how similar activity can move between websites and social platforms.
The pattern to watch is not only whether a single account is banned. The larger issue is whether influence operations keep using generative AI to create political content at scale, test messages across ideological groups, and imitate legitimate publishing outlets.
OpenAI’s action removed one identified cluster. The broader challenge is that tools built for fast writing can also make low-cost manipulation easier to attempt, even when the resulting posts receive little visible engagement.