AI Turns Fake Journalists Into Scalable Propaganda Networks

AI systems are giving influence campaigns a faster way to manufacture the signals of credibility: bylines, profiles, institutions and a steady stream of publishable material.

OpenAI says it disrupted two operations linked to Russia and Iran that used ChatGPT to support covert political messaging. The campaigns took different forms. One built a group of supposed journalists who placed articles in international media. The other operated through a think tank in Latin America, where local workers were reportedly unaware of the Russian connection behind the organization.

The cases show how propaganda networks can use AI at several points in the information chain. ChatGPT can help produce drafts and social media comments, but the larger value comes from connecting those outputs to identities, websites, publications and local staff. The result can look like an ordinary media ecosystem even when its direction comes from a hidden sponsor.

OpenAI described the networks in a report on AI-enabled ‘false front’ operations. A false front is an entity that presents itself as independent while serving the interests of another group. The form can be a news outlet, a research center, a civic organization or a professional account with a convincing biography.

The Iranian operation centered on seven journalist profiles. Each profile claimed to be based in a Western country, including the United States. OpenAI said the personas were used to create long-form articles and place them in media outlets around the world.

The company identified almost 100 articles published or syndicated under the operation’s bylines across roughly a dozen online outlets. Those outlets were generally small to medium-sized publications covering international affairs, geopolitics and Middle Eastern events. The earliest article identified in the report appeared in July 2025. Activity increased after the US-Iran war began earlier in 2026, according to OpenAI.

The operation also generated social media comments connected to the war and other related subjects. That gave the campaign a way to extend the same themes beyond published articles and into public discussions.

The Russian operation had a different structure. Its operators used a think tank called the Social Research Center, or SRC, to promote messages aimed at damaging Ukraine’s reputation in Latin America and influencing local politics. OpenAI said Russian actors recruited or co-opted workers in the region, and that those employees did not know they were working for a Russian group.

OpenAI called the SRC the most complex front identity it had disrupted in the previous two and a half years. The available evidence suggested that the organization produced much of its original material through its local staff. That detail matters because the network’s visible output may have involved real people doing ordinary work, with the hidden political direction separated from the people creating and distributing the content.

OpenAI says it disrupted Russian and Iranian influence operations that used ChatGPT to support false-front entities: a network of supposed journalists and a think tank targeting Latin America. The cases illustrate that AI’s main advantage is not simply producing convincing sentences, but helping covert campaigns maintain the identities, institutions and publishing volume that make political messaging appear independent. Effective detection therefore requires examining the network around the content, including biographies, ownership, funding, employee awareness, account behavior and links among supposedly separate outlets.

A fake journalist profile works because it carries several familiar signals at once. It has a name, a biography, a photo, a social media account and a record of published work. Each signal may be weak on its own. Together, they can persuade an editor or reader that a real professional stands behind the material.

One profile identified by OpenAI used the name Ervin B. Hoskins and described the person as an American freelance writer. Accounts using that identity appeared on X and Instagram. Information displayed by the platforms connected the X account to a West Asia Android application and indicated that the Instagram account was based in Iran, according to screenshots included by OpenAI. Both accounts appeared to have been suspended.

The network’s scale came from repeating this model across several identities. A single operator could maintain multiple bylines, tailor articles to different outlets and produce comments for social platforms without building a separate human team for every task. ChatGPT could help with drafting, translation, editing and variations on the same argument.

That does not mean every article under one of the bylines was generated entirely by a chatbot. OpenAI’s findings establish the use of its accounts and the apparent connection to the influence operation. Individual articles may have involved human writers, editors or translators. The important finding concerns the system around the content: a concealed sponsor could create or manage identities designed to appear independent and then use those identities to reach real audiences.

Small outlets can be especially useful in this model. They may have limited editorial staff, tight deadlines and fewer resources for checking a contributor’s history. A polished pitch from a writer who appears to have a clear specialty can be attractive when a publication needs material about a fast-moving international story. Once an article is published, the campaign gains a link that can be reused as evidence of the writer’s legitimacy.

The SRC case highlights a different weakness. A campaign does not need every visible participant to be part of the deception. It can place the hidden work behind a legitimate-looking organization, hire local personnel and allow those employees to carry out tasks that appear routine.

OpenAI said the Latin American staff members were unwitting. Their involvement could give the organization local knowledge, language skills and professional relationships. It could also create original material that felt less like translated foreign messaging. Articles and reports produced by people who understand the local political environment may appear more grounded to readers than content released directly by a foreign state actor.

The think tank format adds another layer of authority. Research centers routinely publish reports, organize events, comment on elections and offer analysis to journalists. Their branding gives political claims a research-oriented setting. A reader encountering an SRC report might focus on its apparent evidence and authorship while missing the question of who funds or directs the institution.

This structure makes attribution harder. Investigators have to trace account activity, payment or employment relationships, technical links, editorial patterns and the movement of messages across platforms. A visible employee may be genuine. A report may contain original work. The organization itself may still function as a vehicle for a covert foreign campaign.

The model also shows why AI is only one component of the threat. ChatGPT may help an operation produce text or organize information, but human recruitment, institutional branding and distribution relationships give that text a path into the public sphere. The network succeeds through coordination across those layers.

The next propaganda breakthrough is not better writing; it is cheaper credibility.

AI matters because it lets hidden sponsors maintain the bylines, institutions and constant output that make influence campaigns look like ordinary public life. The most dangerous content may be polished by humans and published by real employees, while the deception remains in the network behind them.

Propaganda has always depended on repetition and presentation. A message gains reach when it appears through several voices, formats and locations. AI lowers the time and labor required to maintain those appearances.

A campaign can use a model to:

The list describes capacity, not proof that every listed technique appeared in these two operations. OpenAI’s report specifically tied the accounts to article production, journalist personas, social media comments and the operation of a false-front think tank.

The practical change comes from volume and speed. A propagandist once needed writers, translators, editors and account managers to keep multiple identities active. A model can assist with much of the repetitive work. Human operators still need to choose targets, approve claims and manage relationships, yet the cost of maintaining a wide network can fall sharply.

That scale creates an investigative problem. One suspicious post may be easy to dismiss. Hundreds of posts spread across different accounts and publications can create the impression of broad agreement. Repetition starts to look like confirmation. A statement cited by several websites may seem independently reported even when the sites received it from the same hidden source.

The most important asset in these campaigns is the gap between what an entity claims to be and who controls it. A journalist persona claims professional independence. A think tank claims research independence. A local employee appears to be participating in a normal workplace. Each claim gives the message a different kind of cover.

AI helps maintain that cover because it can imitate the surface features of professional communication. It can produce clean prose, adapt a tone to a publication and fill in the frequent output expected from an active writer or institution. Readers and editors often judge credibility through those surface features because they cannot investigate every source in depth.

That pressure is especially strong during a war or political crisis. News moves quickly, audiences search for explanations and outlets want informed commentary. The Iranian operation reportedly increased its activity after the US-Iran war began. A campaign that can deliver plausible analysis during a period of confusion has a better chance of shaping the first version of events that reaches an audience.

A false identity also gives operators room to change tactics. A journalist can publish a long analysis, comment on another outlet’s article and promote a related claim on social media. A think tank can issue research, invite attention from journalists and build contacts with local organizations. The same political objective can travel through formats that appear unrelated.

Readers and platforms often look for signs of machine-generated language, such as repetitive phrasing or generic structure. Those clues can help, but they address only one part of the operation. Human-edited text can pass as ordinary writing, and a genuine employee can produce material for a compromised institution.

Investigators need to examine the network around the words. Useful questions include:

Newsrooms can add basic source checks to their contributor process. A byline search should include older work, verifiable contact details and consistent expertise. Editors can ask for reporting notes, source lists and disclosure of affiliations. A publication does not need to assume bad faith from every unfamiliar writer. It does need a process that makes invented identities harder to use.

Platforms face a related challenge. Removing one account after a campaign has already built a publishing history may have limited impact. Detection works better when platforms share signals across services and examine coordinated behavior. A suspended profile on one platform may be connected to an active account elsewhere through the same name, image, device or posting schedule.

OpenAI’s action also points to the role of model providers. Account bans can disrupt access to a tool, especially when a network depends on many automated tasks. Providers can publish detailed threat reports, preserve relevant evidence for researchers and coordinate with platforms and investigators. Those steps will not remove propaganda from the internet, but they can expose the infrastructure that supports it.

Readers often encounter an article after it has passed through several layers of distribution. A post may quote a report, which cites a think tank, which relies on a journalist whose identity was created for the campaign. Each step can make the original claim appear more established.

Source awareness means checking the path behind the claim. A publication’s name alone provides little information about its ownership or editorial standards. A writer’s profile may show activity without proving a real background. An institution’s title may sound neutral while its funding and direction remain hidden.

The useful habit is simple: follow the source one step deeper. Check who wrote the material, where else that person has published, who operates the outlet and whether independent reporting confirms the central claim. During a fast-moving conflict, that process may take longer than sharing a post. It also reduces the chance that a manufactured network will turn uncertainty into apparent consensus.

The two operations described by OpenAI show where the pressure is heading. AI gives covert campaigns efficient production, while fake journalists, think tanks and local personnel provide social cover. The technology matters because it expands the network. The credibility damage comes from the institutions and identities that make the network look ordinary.