In a world where the line between reality and carefully manufactured fiction keeps getting harder to see, OpenAI has quietly pulled back the curtain on two of the more sophisticated influence operations to surface in recent years. The company, best known for building ChatGPT, announced on October 12 that it had disrupted covert campaigns originating in Russia and Iran, both of which used its own AI tools to help spread deceptive political narratives across the internet. The report is equal parts fascinating and troubling, because it shows that the technological barrier to creating convincing propaganda has never been lower. These weren’t clumsy operations full of typos and obvious photo-shopping. They were layered, patient, and often convincing enough to get real articles published in real media outlets, to deceive actual employees into doing work for the wrong people, and to plant false documents that triggered real official responses. The whole thing reads less like a spy thriller and more like a grim preview of what the information battleground looks like when artificial intelligence becomes an everyday weapon.
The Iranian operation, which OpenAI codenamed “Bogus Bylines,” was built around a surprisingly simple but effective trick: invent seven fictional journalists and try to pass their work off as legitimate reporting. The operators behind the scheme created detailed fake personas for these nonexistent writers, complete with the kind of credentials and polish that make an author bio look harmless and credible. Then they used those personas to pitch articles about the conflict between the United States and Iran to online publications around the world. According to OpenAI’s investigation, nearly a hundred articles were published or syndicated across roughly a dozen outlets between July 2025 and October 2026, which means the operation ran for well over a year and managed to get real editorial pipelines to bite again and again. What makes it particularly unsettling is how ChatGPT was used to make the content fit seamlessly into each publication’s guidelines. The operators fed submission requirements into the model and asked it to refine draft articles, adjust tone, and produce ready-to-send emails to editors. In other words, the AI wasn’t just generating propaganda; it was acting as a production assistant that knew how to speak the language of the media industry. The articles themselves consistently portrayed the United States as the aggressor and Iran as a resilient victim of international pressure, a familiar narrative in Iranian state messaging but one that now arrived on the desks of editors wearing the clothing of independent journalism. Alongside the articles, the same operators used ChatGPT to generate batches of social media comments designed to amplify the stories and create the illusion of grassroots discussion. The comments rarely gained much traction, which OpenAI rated as a Category 2 activity on its six-point influence operations scale, while the article-publishing effort itself was rated Category 4, reflecting its more meaningful reach into legitimate news ecosystems. The real audience, it seems, was never going to be the tweeters; it was the people who read an article on a news site and trusted it simply because it looked like journalism.
The Russian operation, dubbed “Dark Clark,” took a different and arguably more devious route by exploiting something even more valuable than fake journalists: real people. At the center of this campaign stood a supposed Latin American research organization called the Social Research Center, or SRC, which presented itself as an independent academic and policy institute. But according to OpenAI, the entire operation appeared to be controlled by a fabricated persona known as “Mia Clark,” a woman who supposedly ran the center and coordinated its research. What’s chilling is that local employees were apparently hired to conduct interviews, gather data, and produce research reports, all while having no idea that the organization they were working for was a front for Russian influence activity. These staff members likely believed they were doing meaningful work for a legitimate think tank. Their bylines lent the operation an authenticity that no fake persona could ever manufacture on its own, because real people with real professional reputations were attached to the output. OpenAI identified more than sixty articles on the SRC website, most of which appeared to be original content, not recycled talking points. The intended purpose of all this research was to undermine Ukraine’s reputation and shape domestic political opinions in certain Latin American countries, particularly Argentina and Bolivia. By positioning the SRC as a neutral academic voice from Latin America, the Russian operators could make arguments that would look suspicious if they came from official Russian sources, but seemed fresh and believable when attributed to local researchers. It’s a classic influence operation technique, layering the operation in multiple disguises, but the fact that real employees were unknowingly involved adds a tragic dimension. They weren’t just pawns; they were unwitting authors of the very deception they were being used to carry out.
Dark Clark didn’t stop at research reports. OpenAI’s investigation uncovered a whole toolkit of manufactured evidence, including fabricated documents and even audio scripts, all of which were assisted by ChatGPT. In one particularly brazen example, the operators drafted a fake letter that implicated Ukraine’s honorary consul in Panama in corruption. It wasn’t enough to simply invent the letter; they also created a matching printed document and a voiceover that went along with it, giving the conspiracy a kind of multimedia life. That fabricated package was posted on TikTok through an account that pretended to be a news outlet, blurring the line between user-generated content and actual reporting. Once the video entered circulation, it was hard for casual viewers to tell whether it was genuine or not, and that was precisely the point. The campaign also targeted schools in Peru and Ecuador with fabricated communications, apparently trying to create confusion or fear within educational communities. What makes this even harder to untangle is that OpenAI found evidence that some of the operation’s claims were picked up by real media coverage and prompted official responses from local governments. But the company was careful to add a dose of reality to its findings. Internal reports from the operation sometimes exaggerated how well the campaign was working, taking credit for events that probably had nothing to do with it or inflating the reach of content that few people ever saw. This is an important reminder that even the people running these operations don’t always have a clear picture of what they’re accomplishing. They may be skilled at creating deception, but they’re just as vulnerable to wishful thinking and delusions of impact as anyone else. In the middle of all that, there were real institutions reacting to fictional documents, real students or parents reading fake communications, and real journalists covering stories that were born as synthetic fabrication.
OpenAI’s assessment of Dark Clark was especially notable because the company gave it a Category 5 rating on its six-point influence operations scale, the highest rating the company has assigned to any operation since it began issuing these reports. That’s a significant admission. On one hand, the operation shows how AI can be used to streamline the grunt work of influence campaigns, translating content, polishing language, generating variations of the same message, and helping operators move quickly across different languages and formats. On the other hand, OpenAI was also careful to note that both the Iranian and Russian campaigns were, in many ways, extensions of influence operations that existed long before generative AI. The technology made them easier to run, but it didn’t create the core strategy out of thin air. Fabricating documents, creating fake personas, setting up front organizations, and planting stories through credible-looking channels are all old tricks. What has changed is the scale, speed, and accessibility of the tools needed to pull them off. A task that once required a team of writers, translators, editors, and social media specialists can now be handled by a handful of operators armed with a subscription and a good prompt. OpenAI said it had banned the accounts involved and shared its findings with relevant authorities, but the broader implications are harder to contain. The report makes clear that ChatGPT isn’t just a tool for writing essays, it has been absorbed into the existing playbook of state-backed influence operations, and there is no simple way to police that kind of misuse without limiting access for everyone.
The most unsettling part of the entire report isn’t the scale of the deception; it’s how quiet and ordinary it all looks from the outside. The Iranian articles appeared in established publications, likely because they met editorial standards and read like real journalism. The Russian research center had a public-facing website, real research, and employees who probably thought they were fighting for truth and transparency rather than quietly serving a foreign influence operation. For readers, that means even a well-sourced article with a real author byline could be part of a coordinated campaign against a government or an international hospital system. In an age where information flies across borders in milliseconds, the report is a powerful reminder that legitimate outlets and respected institutions can be unintentionally turned into vehicles for propaganda. OpenAI’s investigation offers some comfort because it shows that platforms are paying attention and building systems to spot these patterns. But it also leaves a gnawing question: How many other operations are still running right now, right under our noses, with people on the ground who don’t know who they’re really working for? The answer, almost certainly, is that the two campaigns exposed in this report are only the ones smart enough or visible enough to be caught. Behind the polished reports and refined articles are the same human vulnerabilities that have always fueled propaganda: trust in familiar names, respect for research institutions, and the instinct to believe something looks credible when it arrives with the right styling. As AI continues to make deception cheaper and easier, that instinct is no longer enough to protect us. We need to think more carefully, verify more diligently, and understand that in the world where algorithms write and people post, the most dangerous fiction is often the one that looks exactly like the truth.