Iranian campaign planted fake articles in real U.S. publications using ChatGPT

About 100 AI-generated articles were published across at least 20 news outlets, OpenAI said.

By Miriam Waldvogel

“The bill will follow Americans beyond the fighting itself,” an opinion piece in the Port St. Joe Star, a tiny Florida weekly newspaper, warned of the Iran war last month.

The article ran under the byline Ervin Hoskins, which has also appeared on articles critical of the war in larger publications including Daily Kos and Middle East Monitor, which together have more than 3 million followers on Facebook.

But ChatGPT maker OpenAI said Thursday that Hoskins’s purported work was generated using its chatbot by users based in Iran as part of an “influence campaign.” The campaign placed about 100 fake articles across at least 20 news outlets, the company said.

The author photo on the Hoskins column published by the Star was generated using OpenAI’s artificial intelligence tools, according to the company’s online verification tool.

As artificial intelligence tools have become more powerful, some news outlets have been forced to delete AI-generated content published under fake bylines that initially went undetected by editors. Business Insider and Wired removed a series of articles last year that appeared under the byline Margaux Blanchard and appeared to be AI-generated.

But while it was unclear who or what was behind the stories written by Blanchard and some other apparently fake personas, OpenAI said in a Thursday report on use of ChatGPT in political influence campaigns that Hoskins was part of an operation focused on seeding content about U.S. military action in Iran. Many of the articles highlighted negative impacts of the war on Americans.

OpenAI said that Hoskins was among seven bylines that appeared on the series of AI-generated articles, made by users who sent instructions to the chatbot in Persian. The operations appeared to be run by a commercial outfit as a “for-hire influence campaign,” the company said. (The Washington Post has a content partnership with OpenAI).

David Adlerstein, publisher of the Port St. Joe Star, said in an email that the Hoskins column had caught his eye after it was pitched to him by a PeaceVoice, a small liberal think tank in Oregon, that had been sending the Star content for “years.” Gulf County, where Port St. Joe is located, has a population of about 15,000.

“Because we’re in a very red part of Florida, I usually don’t run PeaceVoice columns, but this one struck me as thought-provoking and not a strident attack on Mr. Trump, which would be problematic for much of our readership,” Adlerstein said. Due to his existing relationship with PeaceVoice, he did not feel a need to verify the origin or the column or the author’s identity, he said.

Tom Hastings, the founding director of PeaceVoice, said the organization beefed up its existing no-AI policy after OpenAI contacted him.

OpenAI said it had banned the accounts behind the campaign and that it blocks access to ChatGPT from Iran.

“The overall activity appeared consistent with a commercial actor running a for-hire influence campaign, but we are unable to identify the particular actor involved,” the company said in its report. “We have shared information on this case with the relevant authorities.”

Middle East Monitor and Daily Kos said they had taken down the articles written by the fake bylines.

The Daily Kos articles by Hoskins appeared on a community platform not subject to review by the outlet’s staff before publication. Markos Moulitsas, the site’s founder, said in an email that such a campaign was “incredibly corrosive” to the platform and that he was examining how to prevent future incidents.

“We already struggle with community members accusing others of being bots or fake accounts simply because they disagree with them. An operation like this pours gasoline on that fire,” he said. “The very trust of our community is at stake.”

Articles written by authors OpenAI said were part of the Iranian campaign covered a range of topics in the Middle East but were often critical of the Iran war’s effects on Americans. Many of the outlets were focused on policy and foreign affairs.

After generating articles in English, the ChatGPT users in Iran prompted the chatbot to adjust the draft to fit submission criteria for a specific outlet and to generate a pitch email to send to an editor there.

The purported authors of the articles were described as freelance writers, policy analysts and peace advocates. The Middle East Monitor, which brands itself as supportive of the Palestinian cause, had published at least 21 stories under bylines OpenAI said were fake including Hoskins, Michael Harrison, Jenny Williams and Alice Johnson.

Some of the bylines, including Hoskins, also appear on social media accounts, and in one case a profile on blogging platform Medium, that shared a name and profile image with published articles. An email account apparently used to send the Hoskins article to PeaceVoice, and a LinkedIn account in his name with a photo that appeared with some bylines, did not respond to a request for comment.

At least one other purported author, Ericka Feusier, used an AI-generated profile picture, according to OpenAI’s tool that verifies whether images were generated using the company’s technology.

OpenAI’s report on influence campaigns mounted using its technology also described a cluster of ChatGPT accounts in Russia that were banned after they were used to operate what appeared to be a fake think tank in Latin America. OpenAI said there was evidence the project may have co-opted real humans to work for the organization.

The AI company said that it published the research to inform regulators, other companies and the public of “how we see threat actors trying to leverage AI, and to shine a light on the vulnerabilities they try to exploit.”

The outlets impacted by the campaign vary in reach. But while they appear to have published AI-generated articles under fake bylines, real people have read them.

Hoskins’s column in the Star prompted a letter to the editor — a rare event for an opinion piece about a national topic, Adlerstein said. He added that he is now looking into introducing an explicit AI policy.

Hastings, of PeaceVoice, said the article appeared to be genuine when he saw it.

“I found nothing remarkably pro-Iran in his work, which I would never publish, just a legitimate critique of Trump’s war of choice. If Mr Hoskins sends in another I will simply require that we meet by Zoom to discuss,” he wrote in an email. “At least for now, that’s the best I can do.”