Every article has a voice. It has a name, a face, a biography, and sometimes a social media presence. You read it and think you are listening to a real person with real opinions, perhaps an American journalist living in New York or a freelance writer with a particular outlook on world affairs. But sometimes the person behind the words does not exist at all, and the words themselves were generated by artificial intelligence, not human experience. That is the unsettling reality behind a recently disclosed influence operation. According to findings released by OpenAI and reported by The Washington Post, Iran-based operators used OpenAI’s tools to prepare nearly 100 articles that were published or redistributed across more than a dozen websites under fictitious journalists’ names. The operation dates back to July 2025, and it was designed to shape English-language readers’ views on the United States, Iran, Israel, and the wider Middle East. OpenAI said it blocked the accounts linked to the operation, and noted that the users had hidden their locations with virtual private networks. But the incident raises important questions about how easily artificial intelligence can be used to create convincing fake voices in public conversations.
The mechanics of the operation were careful, methodical, and deeply human in their detail. The operators gave ChatGPT instructions in Persian, asking it to polish English-language articles about the US-Iran conflict, adapt those articles to the specific submission requirements of different publications, and draft messages to editors. In other words, they were not just using AI to write content; they were using AI to navigate the professional customs of journalism, from pitch emails to formatting preferences. They also asked the tool to create biographies and support the identities of their purported freelance writers. In total, OpenAI identified seven Western personas, mainly presented as Americans, built through this process. These weren’t crude pseudonyms. They were constructed identities, complete with personal backstories, worldviews, and in some cases social media accounts. One persona, “Erwin Hawkins,” was described as an American freelance writer and anti-war activist. Another, “Michael Harrison,” had operators request a full, polished biography. The effort suggests a level of patience and dedication that resembles traditional espionage or public relations work, except that the intellectual labor was being outsourced to a machine. The goal was not simply to publish information, but to establish believable characters who could move through the world of online media without raising suspicion.
The content of the articles was also more sophisticated than straightforward propaganda. These were not screaming headlines or obvious fake news. Instead, the pieces covered subjects ranging from Israeli politics to US foreign policy, and their arguments were often nuanced. They were not uniformly anti-Israel, which made them harder to dismiss as a simple ideological campaign. One fictitious writer, “Noah Lamington,” examined Israel’s ultra-Orthodox military draft dispute and the stability of Prime Minister Benjamin Netanyahu’s government in International Policy Digest. Another article, written under the persona “Hoskins,” discussed France’s entry ban on Itamar Ben Gvir in Modern Diplomacy. That piece criticized Ben Gvir and even questioned Washington’s response, but it also expressed support for Israel’s security. This mix of positions made the content feel balanced, even thoughtful. An article in LA Progressive questioned whether Jared Kushner’s business ties in the Gulf had influenced US policy toward Iran, a topic that resonates with genuine debates in American foreign policy. These were not fringe conspiracy theories; they were arguments that could sit comfortably alongside legitimate commentary in the opinion sections of small magazines. The authors may have been ghosts, but their words were designed to pass as reasonable, informed perspectives. That is what made the operation so concerning: it blurred the line between authentic opinion and manufactured influence.
The targeted publications were largely outside the mainstream media, but that does not mean they were insignificant. Some had substantial online audiences and trusted readerships within activist and policy communities. London-based Middle East Monitor, which has nearly two million Facebook followers, published at least twenty-one articles under four bylines that OpenAI identified as fictitious, according to the report. After being contacted, the outlet said it removed the articles identified as AI-generated and was reviewing its opinion submission procedures. Material also appeared in Daily Kos’s community section, where posts are not reviewed by the site’s editors, allowing the fabricated pieces to slip through without the same level of editorial scrutiny. Publication on established platforms, even smaller ones, gave the operation a veneer of legitimacy. Once an article appears on a website with a familiar name, readers are more likely to trust it, share it, and absorb its arguments. The network sometimes followed publication with social media promotion, amplifying the voices of fake writers and helping their ideas reach even larger audiences. This distribution strategy reflects an understanding of the modern information ecosystem: brand names and bylines still matter, even in an age where content is shared and reshared at lightning speed. The operators were not trying to break into the New York Times opinion section; they were looking for cracks in the system where well-crafted, AI-assisted articles could pass as legitimate and then spread organically.
The operation extended far beyond articles. OpenAI said the network also used ChatGPT to generate social media comments that portrayed the United States as aggressive, Israel as an unreliable ally, and Iran as a victim resisting pressure. Some of those comments even promoted the articles the operators had placed themselves, creating a feedback loop where fake content was boosted by fake engagement. This is a particularly insidious tactic, because it manufactures the appearance of grassroots support or outrage, making a fringe position seem more popular than it really is. The network also targeted Iran International, a Persian-language broadcaster critical of the Iranian government. According to OpenAI, the operators used ChatGPT to produce dozens of batches of hostile replies to the outlet’s posts. This suggests the campaign was not only trying to influence Western readers, but also to harass and discredit a media organization that represents a political threat to the Iranian establishment. One article published under the name “Sofia Gonzalez” in LA Progressive and CounterPunch alleged that Netanyahu pushed Trump into war and accused Iran International of helping create the impression that Iranians welcomed the attacks. This was a direct attack on the credibility of a broadcaster that many Iranians rely on for independent news. The operation was therefore multi-pronged: it sought to shape foreign opinion, manipulate social media, undermine a specific media outlet, and damage the reputation of those who challenge Tehran’s narrative.
The broader implications of this discovery are sobering. Artificial intelligence is no longer just a tool for automating tasks or generating text; it is becoming a weapon in the battle for perception and narrative. The Iranian operation, as described, was not a massive state-sponsored propaganda machine in the traditional sense. It was a relatively small, agile effort that used advanced language models to multiply its reach and create believable fictional characters. This is a democratization of influence operations, in a dangerous sense. Actors who once needed large teams of skilled writers, translators, and social media managers can now achieve similar effects with a handful of people and an OpenAI account. The technology lowers the cost of deception, and it allows operators to adapt quickly to the requirements of different platforms and publications. For readers, the challenge becomes even more difficult. We are used to judging the credibility of an article by the author’s name, the publication’s reputation, and the quality of the argument. But when all three can be manufactured, those signals lose their meaning. The incident is a reminder that online information literacy must evolve. We need more rigorous systems for detecting AI-generated content, verifying identities, and holding platforms accountable for what they publish. But we also need something simpler: a healthy dose of skepticism, not only toward what we read, but toward the very idea of authorship in the digital age. The ghost writers in this operation were not human, and their words were not born of lived experience or sincere belief. They were simulations of opinion, designed to pass unnoticed and shape the way we see the world. Recognizing that possibility is the first step toward defending the public conversation from those who would manipulate it.