AI-Generated Bylines: Iran Campaign Placed 100 Articles in Media

A fake writer in a real newspaper

On September 22, the Port St. Joe Star, a small weekly paper in Florida, ran an opinion column titled "The cost of American power in the Iran War." The byline was Ervin B. Hoskins.17 Hoskins does not exist. OpenAI has since identified him as one of seven invented journalists in an influence operation run from Iran. The company named the operation "Bogus Bylines."177 Its operators used ChatGPT to polish English-language articles, fit them to each outlet's submission rules and write pitch emails to editors.3 The author photo on the Star column was not real either. OpenAI's verification tool flagged it as AI-generated.1419

The fake writers were not only published in tiny papers. Bylines tied to the campaign also appeared in Daily Kos and in Middle East Monitor, which have more than 3 million Facebook followers between them.11 OpenAI announced on October 8 that it had banned the accounts involved and passed what it found to authorities.419

Most of the coverage treats this as a story about foreign propaganda. It is also a story about AI content production. The operation reads like an ordinary content shop: articles drafted at scale, pitches tailored to each editor, a generated headshot, and fake reader comments to make it all look popular. Nearly every piece of that work was done with tools anyone can buy.

How the operation worked

The basic account is consistent across outlets. Operators wrote their prompts mostly in Persian, produced content in Persian and English, and used VPNs to hide where they were.7 The seven personas were Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams and Alice Johnson.7 They presented themselves as freelance journalists, policy analysts and peace advocates. Some had matching social media profiles that reused the same names or images.19

The headlines show what the campaign was selling. They included "Is America's Iran Policy Being Privatized? The Jared Kushner Question" and "The Iran War and the Collapse of Western Moral Credibility."12 The Kushner piece ran in LA Progressive, a Los Angeles outlet.13 The earliest article OpenAI found dates to July 2025. The pace picked up sharply after the U.S.-Iran war began this year.10

The content operation had two parts. The first was long-form articles. The second was batches of social media comments, written to cast the U.S. as the aggressor, Israel as an unreliable ally and Iran as a defiant but resilient victim.12 After the articles were published, operators asked ChatGPT for comments they could post as if they came from ordinary readers.18 Some batches praised the campaign's own articles to make them look more popular. Others were more than two dozen sets of hostile Persian-language replies aimed at the UK-based broadcaster Iran International.37

The headshot shows how cheap a fake identity has become

The AI-generated author photo is the detail that should worry the image-generation industry most. A fake byline used to need a stolen photo, which a reverse image search could catch. Now a fake persona can come with a unique, believable face that has never appeared anywhere before. OpenAI confirmed the Hoskins photo was synthetic by running it through its own verification tool.14 That suggests a pattern: the companies that make generative image tools are also becoming the main checkers of their output. Detection works best when it is built into the system that made the image.

The Hoskins persona was not caught by any test of the writing. It was caught by the mismatch between the persona's story and the data platforms collect. Hoskins described himself as an American freelance writer. X's transparency data showed the account connecting through a "West Asia android app," and Instagram listed the account as based in Iran.6 A Reddit commenter in a journalism forum made a similar point. AI-text detectors mostly guess on short news copy, the commenter wrote, and false positives would hit honest freelancers first. Checking who actually filed a piece is the dull fix that works.14 In this case, that argument holds up. The campaign was exposed by identity checks, not by analyzing its prose.

Fake reader comments failed, real outlets worked

For anyone working in AI-generated UGC or ad creative, the main lesson is about what failed. The fake comments went almost nowhere. OpenAI found that the replies usually drew single- or double-digit engagement and were a minority of the real comments under the posts they targeted.3 OpenAI rated the comment effort Category 2 on its Breakout Scale, and the article placements Category 4.4 The operators' own internal reports counted views on the posts they replied to, not views of their own replies. OpenAI said that method greatly exaggerated how well the campaign was doing.7

The articles reached far more people because real publishers lent them credibility and an audience. Some outlets promoted the planted pieces on their own social accounts, reaching people who would ignore anything labeled as Iranian messaging.8 One outlet that ran the articles had almost 2 million Facebook followers, about 355,000 on X and more than 544,000 on Instagram as of August 2026.17 Jessica Brandt, who ran the U.S. intelligence community's Foreign Malign Influence Center from 2023 to 2025, told CNN that AI is helping foreign adversaries hide their role and "launder their content through authentic channels."15

The takeaway is that cheap AI generation does not create reach. The scarce thing is a channel that readers already trust. AI is very good at producing content that a trusted gatekeeper will accept: clean prose, the right length and a pitch tailored to each editor.

Where the reporting disagrees

The coverage agrees on how the operation worked but splits on three points.

Scale. OpenAI's report counts almost 100 articles across about a dozen outlets.6 The Washington Post reported more than 100 articles in at least 20 outlets.11 The Daily Wire noted the gap directly.17 The likeliest explanation is that the Post did its own reporting and found syndication and reposts beyond OpenAI's original list. If so, OpenAI's count should be read as a minimum.

Who did it. Headlines from CNN and others refer to "Iranian operatives."12 OpenAI itself was more careful. It said the activity looked like a commercial firm running an influence campaign for hire, and that it could not name the firm or say whether it had ties to the Iranian government.1213 That matters. A for-hire firm is essentially a disinformation contractor selling a content service, which is a different and more mobile threat than a single state agency.

How big a deal it is. The report also covered a Russian operation called "Dark Clark." OpenAI rated it Category 5, the first time it has given that rating to an operation it disrupted.2 Security analyst Lukasz Olejnik argued that even under OpenAI's own framework the rating is too high.1 OpenAI's broader conclusion is more modest than the headlines. It said the AI tools made the groups faster and more fluent, but their methods mostly looked like influence operations from before generative AI.4 OpenAI compares the Iranian personas to "Alice Donovan," a front for Russian military intelligence whose articles appeared in Western outlets in 2016 and 2017.1

The wider AI propaganda picture

Bogus Bylines did not happen in isolation. Since the war began, young Iranians have made viral AI-generated Lego-style videos mocking U.S. leaders, and President Donald Trump has posted AI-generated videos showing U.S. strikes on Iran.16 CNN also reported that a false, AI-assisted U.S. intelligence report about a Chinese ship carrying nuclear components nearly led the U.S. military to board the vessel.20 AI-made media is now part of the fighting, on several sides and in several formats.

Publishers are starting to respond. David Adlerstein, publisher of the Port St. Joe Star, said the column came through PeaceVoice, a group that had supplied commentary to his paper for years. Because of that relationship, he did not check the author's identity himself. He is now considering a written policy on AI-generated submissions.19 PeaceVoice has tightened its rule against AI-written submissions, and Middle East Monitor has taken down the articles.19 A Middle East Monitor representative said the publication is reviewing how it handles submissions.12

Analysis: verify the person, not just the text

The takeaway from Bogus Bylines is not that AI writing is impossible to spot. It is that small and mid-sized publishers relied on trust where they should have checked identities. Many depend on unpaid contributors and syndication partners to fill space, and an AI-assisted pipeline of fake writers can exploit that cheaply.5 The tools that make content production efficient for legitimate marketers and creators also made this operation efficient.

The practical fixes are about the person, not the prose. Publishers can confirm that a contributor is real, has a verifiable history, and has a face that was not generated. AI companies can keep building detection into their image tools and keep watching for coordinated behavior across accounts, which is how OpenAI says it caught this one.9 Synthetic text and synthetic headshots will keep getting more convincing. Whether a real person stands behind a byline is still something an editor can check, as long as they do.