Fake Journalists Used ChatGPT to Enter Real News Sites
OpenAI says Russia- and Iran-origin influence operations used ChatGPT to refine fake documents, pitches and articles that reached real media outlets.
OpenAI says it disrupted two sophisticated influence operations that used ChatGPT to help disguise propaganda as independent journalism. One originated in Iran and placed nearly 100 articles under false bylines. Another linked to Russia built a fake think tank, produced fabricated leaks and recruited unwitting participants in Latin America.
The finding is important because the campaigns moved beyond low-engagement social-media spam. Their operators persuaded real small and medium-sized outlets to publish or syndicate material, giving deceptive personas access to audiences and institutional credibility they could not easily create alone.
OpenAI called the Iranian campaign “Bogus Bylines.” It used seven journalist identities with Western-sounding names, biographies and social-media profiles. Operators asked ChatGPT to refine English-language articles about the US-Iran conflict, check submission requirements and draft emails pitching editors.
The company found almost 100 published or syndicated articles across roughly a dozen outlets. Some were then promoted through the publications’ own social accounts. The operation also generated batches of comments portraying the United States as an aggressor, Israel as an unreliable ally and Iran as resilient.
The Russian-origin network was assessed as Category 5 on OpenAI’s six-level influence scale, the highest-reach campaign the company says it has disrupted. It appears to have co-opted unwitting people in Latin America to operate a supposed think tank and produced fake documents and audio scripts designed to implicate political figures.
OpenAI says both operations resembled pre-AI influence campaigns. The models reduced time spent editing, translating, tailoring style and writing internal reports; they did not invent the basic strategy. Humans created false identities, contacted editors, operated accounts and chose geopolitical objectives.
That distinction prevents two opposite mistakes. It would be wrong to dismiss the campaigns because AI was only one tool. Faster language production and stylistic adaptation can let a small team operate more personas and target more outlets. It would also be wrong to claim ChatGPT autonomously conducted propaganda or that AI alone persuaded editors.
The phrase “Russian and Iranian regime operations” needs qualification. OpenAI traced account origin and behavioral indicators. For the Iranian cluster, it said activity was consistent with a commercial influence-for-hire actor but could not identify the specific organization. Government direction is possible, but not proven in the published report.
The campaigns exploited an editorial weakness: many outlets need inexpensive international analysis and rely on contributor biographies they cannot deeply verify. A polished pitch, plausible headshot and consistent social profile can pass superficial checks. AI makes that package easier to maintain across multiple identities.
Newsrooms can respond without banning freelance submissions. Editors should verify identity through live calls, institutional email, publication history and reverse-image searches. They should examine whether multiple authors share unusual phrasing, contact patterns or coordinated social promotion. Payment information can also expose networks of supposedly unrelated contributors.
OpenAI banned the relevant accounts and shared information with authorities and industry partners. Account removal disrupts a workflow but does not erase already published material or stop operators from moving to another model. Publishers must review affected articles and disclose corrections when authorship was deceptive.
The report also found that social engagement was often weak and internally exaggerated. Operators counted views on original posts as though their own replies caused them. That suggests influence contractors may deceive their sponsors as well as audiences.
OpenAI’s findings also complicate the popular image of automated propaganda. The operators did not simply publish raw chatbot text at enormous scale. They combined invented identities, human outreach, editorial adaptation and legitimate submission channels. In other words, the model lowered the cost of persuasion work but did not replace the social engineering needed to enter a publication. Newsrooms can respond by verifying unfamiliar contributors, checking biographies and reused profile images, retaining pitch histories and asking for disclosure when a submission depends heavily on generative tools. Detection cannot rely on awkward prose because modern systems can imitate competent professional writing.
The outlets that accepted the material should examine what passed editorial review, inform readers when warranted and correct any false biographical claims. Accountability belongs to both the deceptive operator and the gatekeeping system it exploited.
What to watch next
Watch whether named outlets remove or label the articles, whether governments attribute the operations and whether platforms share indicators across services. The deeper problem is not that AI can write propaganda; it is that real editorial systems can mistake manufactured identities for journalists.