Iran-Based ChatGPT Users Placed 100 Articles in Real Newsrooms
OpenAI says an influence campaign used ChatGPT to write, tailor and pitch op-eds that ran in at least 20 outlets, including a Florida weekly. The weak point was not the model.
The op-ed ran in the Port St. Joe Star, a weekly in Gulf County, Florida, a county of about 15,000 people. It carried the byline Ervin Hoskins. According to OpenAI, Ervin Hoskins was generated with ChatGPT by users in Iran, as part of an influence campaign that placed roughly 100 articles in at least 20 news outlets.
OpenAI disclosed the operation on Thursday in a report on ChatGPT use in political influence campaigns, and The Washington Post, which has a content partnership with OpenAI and said so in its story, reported how the pieces got into print.
The workflow
The process was simple and repeatable. Users writing instructions in Persian had ChatGPT draft articles in English. They then had the chatbot reshape each draft to fit a particular outlet's submission guidelines, and write a pitch email to the editor. The purported authors were presented as freelance writers, policy analysts and peace advocates.
There were seven bylines across the series. Besides Hoskins, the Post named Michael Harrison, Jenny Williams and Alice Johnson, whose pieces appeared at Middle East Monitor, and Ericka Feusier, who used a profile photo that OpenAI's own image-verification tool flagged as AI-generated. Middle East Monitor alone published at least 21 stories under bylines OpenAI said were fake. Hoskins pieces also appeared on Daily Kos's community platform.
The articles focused on the U.S.-Iran war, and many emphasized its costs to Americans. Headlines included "The Jared Kushner Question" and "The Iran War and the Collapse of Western Moral Credibility."
How a fake writer gets past an editor
The Port St. Joe Star did not receive the Hoskins column cold. It came through PeaceVoice, an Oregon organization that distributes opinion pieces and that the paper had worked with for years. Publisher David Adlerstein told the Post he did not independently verify the writer because of that existing relationship. He is now considering an explicit AI policy.
PeaceVoice founder Tom Hastings said the column appeared genuine and that the group strengthened its no-AI policy after OpenAI contacted it. "I found nothing remarkably pro-Iran in his work," he said.
That is the uncomfortable center of the story. The articles did not need to be brilliant. They needed to be competent, on-topic and formatted to each outlet's rules, and to arrive through a channel editors already trusted. A chatbot handles the first three. The fourth is a human vulnerability that predates AI.
Middle East Monitor and Daily Kos both removed the articles and are reviewing their processes. Daily Kos founder Markos Moulitsas called the campaign "incredibly corrosive" to the platform.
Who was behind it
OpenAI stopped short of naming a government. It said it could not identify the specific actor or establish a direct link to the Iranian state, and that "the overall activity appeared consistent with a commercial actor running a for-hire influence campaign." It banned the accounts involved, says it blocks ChatGPT access from Iran, and shared information with relevant authorities. The same report described a separate Russia-linked cluster of ChatGPT accounts that ran an apparently fake Latin American think tank.
This is not OpenAI's first disclosure of this kind. In August 2024 it banned accounts it tied to a Tehran-backed group known as Storm-2035, which generated long-form articles and short social posts about that year's U.S. presidential race. OpenAI said at the time that the campaign did not appear to achieve meaningful audience engagement.
The difference two years later is placement. Storm-2035 mostly published on its own sites and social accounts. This campaign got into real mastheads.
What newsrooms can actually do
The defense is old-fashioned and cheap. A phone call. A video chat. A request for a prior clip or a professional profile that predates the pitch. Verifying that a contributor exists has always been part of running an opinion page; for a small weekly with one or two editors, it is often the step that gets skipped when a trusted syndicator vouches for a piece.
Syndicators are now a pressure point. One trusted intermediary forwarding a fake writer can place the same column in many outlets at once. PeaceVoice's policy change is the right instinct, but a no-AI policy only works if someone checks who the human is.
AI-detection tools are not the answer here. The articles were edited for each outlet, and the problem was never that the prose read as machine-made. The problem was that nobody confirmed a person stood behind the byline.
The takeaway
The model made the operation efficient. It did not make it possible. Op-ed pages have always relied on good faith about who is writing, and a for-hire campaign just priced that good faith at a few prompts per outlet. OpenAI caught this one by watching its own platform. The next one may use a model nobody is watching, and the only reliable check left will be the editor who picks up the phone.