A small newspaper in Florida published an opinion column by Ervin B. Hoskins, a freelance writer, warning Americans about the consequences of the war with Iran. Similar articles bearing his name appeared in other publications, including Daily Kos and Middle East Monitor.

Hoskins, it turned out, was not a freelance journalist at all. He was one of seven fictitious authors in an Iran-based influence operation to place almost 100 articles in online publications around the world.

The scheme was exposed in an OpenAI investigation published on October 8. The perpetrators used ChatGPT to revise English-language articles, adapt them to individual publications’ editorial requirements, and compose e-mails pitching them to editors. Some of the resulting pieces were promoted through the publications’ social media accounts.

The operation did more than manufacture political arguments. It gave them the appearance of independent authorship and, once they were published, some of the credibility of established media outlets.

None of this necessarily meant the arguments were false. An article could make a legitimate criticism of American or Israeli policy while concealing the identity of whoever had commissioned or prepared it. Readers were being misled not necessarily about the claims but about who was making them.

That distinction matters for journalism. Newsrooms devote considerable effort to checking facts and assessing the quality of submissions. They also depend on something less visible: the assumption that a contributor is whom they claim to be.

The OpenAI case shows how that assumption can be exploited. It also adds a new dimension to an Anthropic investigation published in September, which documented several Iran-related operations using Claude. In one, an operator supplied the system with roughly 8,400 Telegram posts by an Iranian activist and instructed it to imitate his writing. The output was used in conversations with people who, according to Anthropic, did not know they were communicating with an AI-assisted impersonation.

Other cases involved state-aligned organizations disguising the institutional origins of political material, and security-linked operators using AI to identify people behind anonymous social media accounts.

The activities were conducted by different actors, including political adversaries, and there is no evidence that they formed part of a single operation. Yet each involved altering the connection between information and the identity presented with it.

The two investigations are particularly revealing because OpenAI and Anthropic could examine how their systems were used while the operations were being prepared. Their researchers saw instructions, datasets, and production processes that journalists working from published material would rarely encounter.

That vantage point does not settle every question about responsibility or impact. It does, however, show why identifying the source of political information is becoming as important as checking its contents.

The Journalists Who Did Not Exist

OpenAI called the Iranian operation Bogus Bylines. The seven “journalists” were Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams, and Alice Johnson. Their biographies generally described Western freelance writers or commentators interested in international affairs, human rights, and conflict. Most purported to be American; Lamington was presented as living in New Zealand.

The operation did not depend on bylines alone. Several identities had supporting social media profiles. Michael Harrison appeared on Medium, X, and Facebook; Hoskins had accounts on X, Facebook, Instagram, and Threads. Transparency information from the latter’s accounts indicated activity originating in Iran, despite his biography's claim that he was an American freelance writer. Meta removed his Instagram account in August.

The operators even asked ChatGPT to write Harrison’s professional biography. These details helped create the appearance of an established writer. An editor checking a contributor’s name might find previous articles, a biographical description, and several apparently corroborating profiles.

The earliest article OpenAI identified appeared in July 2025. Publication accelerated in March after the outbreak of the 2026 war involving Iran, Israel, and the US. The pieces generally concerned the conflict, international politics, and the consequences of American military action.

The operators’ methods were adapted to the conventions of journalism. They submitted English-language drafts to ChatGPT and asked it to assess them against a particular outlet’s submission guidelines. The system suggested revisions, which the operators could approve before requesting a covering e-mail to the editor. This was more than indiscriminate online propaganda. The material was being prepared for the editorial processes of real publications.

One route became visible when the Washington Post examined how a Hoskins column reached the Port St. Joe Star, a weekly newspaper in Florida. Publisher David Adlerstein received the article through PeaceVoice, an organization that had supplied the paper with material for years. He considered the column thoughtful and appropriate for his readers. The established relationship meant that he did not independently verify the supposed author’s identity. PeaceVoice said it had been unaware of the deception and subsequently tightened its rules on AI-generated submissions.

Other publications responded by removing material or reviewing their procedures. Daily Kos noted that the relevant contributions had appeared on its community platform, where users can publish without undergoing the same advance editorial scrutiny as staff writers.

These distinctions are important. A community website and a newspaper receiving articles through a trusted intermediary operate differently. Neither arrangement establishes that editors knowingly participated in the campaign. Indeed, the scheme depended on their not knowing.

OpenAI identified almost 100 articles published or syndicated under the seven names across more than a dozen outlets. Some were subsequently promoted by the publications themselves, giving the operators access to audiences beyond their own accounts.

The apparent authority of the material changed as it passed through those publications. What began as a politically-motivated submission under a fabricated identity became an article carrying the standing of an established outlet.

OpenAI assessed the activity as consistent with a commercial influence operation conducted for hire. It traced the relevant ChatGPT accounts to Iran but could not identify the organization responsible. Any distinction between Iranian origin and direction by the Iranian government remains unresolved. The political content of the articles does not establish who commissioned them.

An Audience That Was Harder to Manufacture

The operation also used ChatGPT to generate large numbers of social media comments, mostly about the US-Iran conflict. Operators supplied posts or screenshots and requested responses in English or Persian. They sought comments portraying the US as the aggressor, criticizing Israel, and challenging American military policy. More than two dozen batches of hostile Persian-language replies targeted Iran International, the London-based broadcaster critical of the Iranian authorities.

Other comments were designed to support the operation’s own articles. The operators fed ChatGPT screenshots of posts promoting those pieces and requested favourable responses, apparently hoping to create the impression of independent public approval.

The results were less impressive than the publication campaign. OpenAI found that identifiable comments generally attracted little genuine engagement. Some received approval mainly from accounts associated with the operation. The operators nevertheless claimed substantial success in internal reports, sometimes comparing the number of views on an original post before and after their replies appeared and treating the difference as evidence of impact. That calculation could not establish whether anyone had seen the comments, much less been persuaded by them.

OpenAI used its six-level influence-operation Breakout Scale to distinguish among the results. The commenting activity received a Category 2 rating, while the article-placement campaign reached Category 4. The difference was not that the published articles had demonstrably changed public opinion. It was that they had entered independent publications with substantial potential audiences. One outlet carrying the material had nearly 2 million Facebook followers, approximately 355,000 followers on X, and more than 544,000 on Instagram in August 2026.

Those figures cannot tell us how many people read a particular article. They do show the distribution network an invented journalist could potentially access once accepted by a publisher. The operators had found something more valuable than a collection of artificial likes: an established media outlet willing to publish their work.

The Russian Connection: Real Writers, Concealed Employers

The second influence operation exposed in OpenAI’s October report followed a different approach. Its operators were based in Russia, and much of their activity targeted Latin America. Their apparent objectives included damaging Ukraine’s reputation and influencing political debates in countries such as Argentina and Bolivia.

OpenAI named the campaign Dark Clark, after Mia Clark, a fictitious identity used to manage a purported research organization called the Social Research Center (SRC). The SRC presented itself as a platform concerned with the Indian diaspora in Latin America and relations between India and countries in the region.

Unlike the Iranian campaign’s invented journalists, many of the people doing its research were real. Internal records examined by OpenAI referred to salaries, staffing, appointments and dismissals, suggesting that the Russian operators exercised control over the SRC rather than merely cooperating with it. The company’s investigators assessed that employees in Latin America were unaware they were working for a Russian-controlled operation.

The work was not necessarily fictitious. The SRC published more than 60 articles, most apparently original. Researchers interviewed specialists, explored economic and political questions and produced substantive material.

This complicates the usual distinction between authentic journalism and fabricated propaganda. A report may be based on genuine interviews and written by a real researcher who has no knowledge of the organization’s concealed management. The deception lies further back, in the institution setting the priorities and controlling the work.

The Russian operators also used ChatGPT to prepare extensive internal reports on their activities. These described attempts to spread false stories about Ukraine and intervene in regional political disputes.

Some claims corresponded to events documented publicly. In Peru, the operation reported an attempt to stir controversy over the 20th-century Ukrainian nationalist Stepan Bandera, using a forged communication encouraging schools to hold events associated with him. The resulting controversy attracted reporting and contributed to discussion involving Ukraine and Poland. Other material involved fabricated claims about Ukrainian military recruitment in Ecuador and attempts to spread misleading information in Bolivia.

Not every achievement claimed in the internal reports could be substantiated. OpenAI found instances in which the operators appeared to take credit for unrelated events or existing political positions. The records therefore exposed two kinds of deception: the attempt to manipulate public information and the attempt to exaggerate the campaign’s success to whoever was overseeing it.

OpenAI rated Dark Clark at Category 5 on its Breakout Scale, the highest rating the company had assigned to an operation it disrupted. Its assessment reflected evidence that some material had penetrated public discussion far enough to elicit fact checks, official denials, or political responses.

The Russian case shows why a real byline does not necessarily settle the question of provenance. A journalist may genuinely have written an article while remaining unaware of the interests controlling the organization that commissioned it.

An Activist’s Borrowed Voice

Where OpenAI found fictitious journalists, Anthropic encountered the imitation of a real political activist.

An operator supplied Claude with around 8,400 Telegram posts written by an Iranian activist and directed an AI agent to adopt his identity and style. The resulting account was then used to conduct live conversations with people who apparently believed they were communicating with the activist himself.

The operator was not simply asking a model to write persuasively. The aim was to make somebody else’s words appear to come from a particular, recognizable person.

The operation also gathered information about people inside Iran. According to Anthropic’s September threat-intelligence report, more than 500 social-media channels were scraped and approximately 51,944 archived messages analyzed. The information was used to create dossiers classifying individuals by city, age, occupation, political alignment and arrest history. Other participants maintained social-media identities, generated avatars and worked with approved source lists, review procedures and persistent AI-agent memory.

Anthropic linked participants to the People’s Mojahedin Organization of Iran (PMOI/MEK) and the National Council of Resistance of Iran (NCRI), which oppose the Islamic Republic. At least four people involved worked for official NCRI media outlets, according to the report.

References to leadership and approval procedures led Anthropic to assess central direction as likely. However, it found neither shared account infrastructure nor visible coordination connecting all the actors. Therefore, it could not establish the extent of centralized control.

The PMOI and NCRI rejected organizational involvement. In a statement issued on September 14, the PMOI said neither organization knew of or was connected to the activities described. It criticized Anthropic for failing to seek its response before publication and argued that responsibility for any wrongdoing by individual members or supporters would rest with those individuals. The organisation said its inquiries had identified one recently closed AI account belonging to a supporter, which it described as distributing publicly-acknowledged positions.

There remains a substantial difference between Anthropic’s evidence of individual involvement and a finding that the NCRI leadership authorised the operation. Employment relationships and indications of approval procedures are relevant; they do not, without further evidence, establish who directed each activity.

The reported impersonation nevertheless illustrates a new difficulty in authenticating political communication. A person’s style, once distinctive enough to help others recognise them, can be reproduced from their public writings. Familiar expressions and characteristic arguments may reassure a recipient who has no way of knowing whether the apparent author produced the message.

Where The Information Came From

Other operations examined by Anthropic involved institutions aligned with the Islamic Republic rather than its opponents. The company identified accounts associated with people working within or on behalf of the Islamic Culture and Communications Organization, the Islamic Propaganda Office of Khorasan Razavi, and the Bina Cultural Observatory of the Islamic Propaganda Organization.

Operators used concepts such as soft war, cognitive warfare, and Jihad al-Tabyin [jihad of clarification] to describe their activities. They employed Claude to prepare campaign plans, develop online personas, and organize information about intended audiences. One project, Manjanegh (Catapult), involved dozens of activists across several Iranian provinces.

According to Anthropic, material originating with Iranian security services was adapted into content for different online identities. The institutional origin was removed before distribution through Iranian and international platforms, where more than 100 channels helped amplify it.

Anthropic called the practice “attribution laundering”. The expression is useful because it identifies a deception that does not necessarily depend on altering the content of a claim. What changes is the apparent speaker. Information originating within a government institution can reappear as the work of an independent commentator. Organized distribution can resemble spontaneous public discussion.

At the Bina Cultural Observatory, Anthropic found an operator generating material in the voice of a spokesman for the Islamic Revolutionary Guard Corps. During the 2026 conflict, the same operation falsely attributed claims to Western organizations including the Center for Strategic and International Studies, Brookings Institution, and RAND. Attaching the name of a recognised research organisation to a statement can lend it an authority the
original source did not possess. A reader encountering the finished item may have little indication that the attribution is false.

The operations attributed to NCRI-associated individuals and those linked to institutions of the Islamic Republic served opposing political interests. Their activities should not be conflated. They do, however, demonstrate how AI can be used to borrow either a person’s voice or an institution’s reputation. In both cases, the identity presented with the information becomes part of the deception.

From Pseudonym to Person

Anthropic also documented activity with a different purpose: discovery of the real identities of people behind online accounts.

Sixteen Claude accounts were associated with two linked units that Anthropic assessed, with high confidence, as connected to Iranian paramilitary domestic-security entities. The report did not publicly identify a specific parent intelligence organization.

One unit, operating across several provinces, claimed to have profiled 6,388 Iranians in a year. Its analysts worked with a network of more than 155,000 tweets. Another unit, based in Qom, used Claude to develop surveillance tools. Both contributed information to a system called Arman, where records could combine national identification numbers, social-media accounts, telephone data, political and religious beliefs and criminal histories.

Anthropic also described the matching of telephone numbers to identities, collection of online identifiers, phishing for national identification numbers, and mass reporting of Telegram accounts. One analysis singled out 39 opposition and diaspora accounts for surveillance.

For an activist operating under a pseudonym, the danger is not simply that a system might reproduce their writing. Fragments of information scattered across platforms can be combined to identify them. The consequences may extend beyond the internet.

In September, British, American and Dutch security agencies warned about malware known as Chosen Brick, which they said had been used against dissidents, activists and journalists since at least 2025. Attackers could approach targets through messaging applications or through impersonation of familiar contacts or technical-support staff. The malware was capable of collecting messages and other personal information that might reveal a person’s location and daily movements.

Chosen Brick was not part of the activity identified by Anthropic, and no common command structure has been established.

Another operation received a more direct public attribution to an Iranian intelligence institution. On March 19, 2026, the US Justice Department announced the seizure of four domains allegedly used by Iran’s Ministry of Intelligence and Security for hacking, intimidation, and psychological operations. The domains included infrastructure presented under the Handala Hack identity. US authorities alleged that it had been used to claim responsibility for cyberattacks, publish stolen personal information, and threaten journalists and opponents of the Iranian government.

The cases differ in their methods and attribution. Together they show how the manipulation of online identity can involve both concealing the operator and exposing the target.

The Older Problem Behind The New Technology

None of these practices began with generative AI. Governments, political movements and intelligence services have long used intermediaries, pseudonyms, and apparently independent institutions to give their messages greater reach or authority.

The Iran Experts Initiative offers a disputed example from an earlier period. In 2023, Semafor and Iran International reported on correspondence involving Iranian Foreign Ministry officials and researchers associated with institutions in Europe and the United States during the nuclear negotiations. The correspondence described efforts by Iranian officials to maintain relationships with academics and monitor their publications and media appearances.

The International Crisis Group, whose employees appeared in the material, rejected the characterization of the initiative as a Foreign Ministry-directed influence operation. It described the network as an informal forum for exchanges with Iranian officials and researchers, maintaining that no government could dictate its analysts’ positions.

One exchange reported by Semafor followed a meeting in Vienna in May 2014. An individual based in Germany who was involved in the initiative proposed that members draft an essay on the nuclear negotiations that, following review by then-Foreign Minister Mohammad Javad Zarif and his team, could be published under the name of a former official. Zarif responded favorably and suggested that opinion pieces be published under the names of
individuals inside and outside Iran.

The influencer subsequently challenged the reporting and questioned the authenticity of his correspondence. Semafor reported that a forensic examination commissioned by Iran International had found no metadata discrepancies indicating that the e-mails were inauthentic.

The exchange is relevant because it concerns the proposed separation of authorship from the name under which an argument would appear. But the correspondence does not establish how many, if any, articles were published through the suggested arrangement. Nor does contact with Iranian officials demonstrate that every participating researcher acted under their direction. Academics may independently advocate diplomacy, oppose sanctions, or criticize Western policy. Agreement with a government’s position is not proof of an undisclosed relationship.

That caution applies equally when foreign experts appear in Iranian media. An interview may be accurate, selectively edited, translated with a different emphasis, or republished under a misleading headline. The circumstances of publication matter more than whether the speaker’s views happen to suit the outlet.

This was an investigative problem before language models existed. AI has made some of the methods easier to reproduce and harder to recognise, but it has not changed what must be proved.

What The Evidence Can Tell Us

OpenAI and Anthropic occupy an unusual position in these investigations. Unlike journalists reconstructing an operation from websites, social media posts and public documents, the companies can examine how their own systems were used. They may see the instructions given to a model, the material supplied for analysis, the generation of an artificial persona, or the revision of an article before it is submitted for publication.

In the OpenAI case, this allowed investigators to connect work on articles under fabricated identities with material appearing in actual publications. In the Anthropic activist case, the instruction to imitate a particular person’s writing provided evidence of intent that could not reliably have been inferred from linguistic similarities alone.

Yet the view remains partial. Neither company necessarily knows what happened outside its systems, who ultimately financed an operation or how an audience responded. That distinction is evident in the Iranian OpenAI investigation. The commenting campaign generated considerable material but little visible engagement. The fictitious journalists were more successful at obtaining publication, although that success says nothing conclusive about their influence on readers’ opinions.

OpenAI’s findings on Dark Clark also demonstrate why internal records need independent scrutiny. Operators who report to a superior may exaggerate their achievements or claim credit for events they did not cause. Evidence of production, distribution and effect answers different questions. So does evidence linking individual operators to organizations.

OpenAI could trace the Bogus Bylines accounts to Iran but could not identify their sponsor. Anthropic found employment relationships and indications of approval procedures in the NCRI-associated activity, while acknowledging that centralised direction remained uncertain. The PMOI and NCRI denied organisational responsibility.

Similar care is required in the cases linked to Iranian state-aligned institutions. Observing how an account was used can establish what its operator attempted, but the chain of authority may extend beyond what the available evidence reveals. The evidentiary picture becomes stronger when investigators can document direct instructions, editorial approval, financial relationships, shared technical infrastructure, or the movement of material from one actor to another before publication. The mere fact that several people advance similar arguments is much weaker evidence. So is the observation that an article serves a particular government’s interests. Especially in polarized political debates, confusion of agreement with coordination can produce allegations that are both unfair and difficult to disprove.

Following The Information

For editors, the Bogus Bylines affair presents an uncomfortable question: what does it take to establish that an author exists? A professional biography may be invented. Social-media profiles can be built around it. Earlier publications may appear to confirm the author’s identity, although they were obtained through the
same deception. The Florida newspaper’s experience shows that even a trusted intermediary can unintentionally carry such material into a newsroom.

A fabricated byline is not the same as legitimate assistance from an AI writing tool. Authors routinely use editing software, translation services and other means of improving their prose. The relevant issue is whether they have misrepresented the person or institution responsible for the work.

Nor is every anonymous contributor suspicious. Anonymity can be essential for journalists, dissidents and sources facing threats. The task is not to demand public identification in every case, but to establish the claims about authorship and provenance on which publication depends.

A useful investigation begins by tracing information back to its earliest identifiable appearance. Where did a claim first appear? Who supplied it? Was the institution behind it named? Did the wording change when it moved between languages or platforms? Was a qualification omitted or a quotation detached from its original context? Did material presented as independent analysis already exist in an institutional document before publication?

Publication histories can reveal how a claim acquires new authority. A statement originating within a state organisation may circulate through an apparently independent commentator, be cited by another outlet and return to its original media environment as evidence of external agreement. That sequence can be significant when it is documented. The political sympathies of the apparent
commentator, on their own, cannot establish it.

The Russian SRC case adds a further complication: even when the writer and the research are genuine, the organization responsible for commissioning the work may have concealed interests. Establishing an author’s identity is therefore only part of establishing the origin of a publication. The central challenge is to recover the history obscured by the finished article.

The Byline as Evidence

OpenAI’s October investigation matters because the fabricated journalists did not remain confined to obscure accounts or unsubmitted drafts. Their work appeared in genuine publications. It carried ordinary bylines and sometimes benefited from the distribution networks of established media outlets.

Anthropic’s investigation shows other ways in which the apparent identity of a message can be manipulated: an activist's voice imitated, a state institution’s role concealed, a Western research organization falsely cited, or an anonymous person’s identity uncovered through the combination of personal data.

The operations are not equivalent, and the evidence for each supports different conclusions. But the OpenAI case brings the issue into the newsroom itself. Editors have always had to decide whether a contributor's arguments are accurate, original and worthy of publication. Increasingly, they may also need to establish whether the author, the professional history and the institution behind a submission are what they appear to be.

Recognizing artificial prose will not answer those questions. A fabricated writer can produce a persuasive article. A real researcher can work for a concealed organization. A message written in a familiar voice may have been produced by somebody else. What matters is the evidence connecting the published words to the people and institutions responsible for them.

Generative AI has made it easier to create political content, imitate individuals and build credible-looking identities around false claims of authorship. The methods used to establish responsibility must now contend with that reality.

The byline above an article is not merely a name. It is a claim about where the words came from. The Bogus Bylines investigation shows why that claim deserves scrutiny in its own right.