OpenAI Disrupts Iranian Influence Campaign Using ChatGPT

The most unsettling detail in OpenAI’s latest takedown report isn’t the fake journalists. It’s where their articles ended up: published, under real-looking bylines, in real outlets, nearly a hundred of them. An Iranian influence campaign dismantled this week used ChatGPT and seven invented personas to plant long-form articles in about a dozen online publications. And here is the part that should worry every newsroom: OpenAI’s own analysis found that operations landing their content in real media outlets, rather than relying on fake social accounts, tend to have the highest potential reach and impact. The bots were noise. The bylines were the weapon. It is the moment ChatGPT disinformation stopped being a bot problem and became an editorial one.

OpenAI banned ChatGPT account clusters from Iran and Russia running “false front” entities, fabricated identities built to launder geopolitical messaging. The Iranian influence campaign, nicknamed “Bogus Bylines,” got inside real media; a Russian operation dubbed “Dark Clark” ran a fake Latin American think tank partly staffed by people who never knew who they really worked for. Findings went to the relevant authorities.

The Iranian Influence Campaign: Seven Journalists Who Never Existed

The Iranian influence campaign ran on a trick so simple it feels almost old-fashioned. Instead of building fake news sites from scratch, the operators invented seven journalists, gave them Western names, and persuaded real publications to run their work. Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams, Alice Johnson, none of them exist. For the Michael Harrison persona, the operators even asked ChatGPT to write the biography itself. The Hoskins persona’s bio claimed he was “an American freelance writer,” backed by X and Instagram accounts.

Those accounts gave the operation away: X’s transparency data showed Hoskins connecting through a “West Asia android app” while Instagram placed him in Iran. Both are now suspended, and this wasn’t the personas’ first exposure. Meta had disrupted an earlier Iranian operation featuring a Michael Harrison persona, banned the Hoskins Instagram in the summer, and three personas’ X accounts were suspended around the same time. The network kept rebuilding; it kept getting caught.

Investigators found almost 100 articles planted by the Iranian influence campaign under fake bylines across roughly a dozen small-to-medium outlets covering international affairs and the Middle East, none mainstream, but some with serious reach, including one with nearly two million Facebook followers. The output stretched over about fifteen months, accelerating after the US-Iran conflict broke out.

Named outlets show how far the material traveled. The Kushner piece ran in the LA Progressive, a Los Angeles publication that describes itself as an independent, justice-driven outlet. NPR reports the operation also placed material in Daily Kos’s community section, where anyone can post without staff pre-screening, and the persona’s posts were still up. The London-based Middle East Monitor also ran the operation’s articles; its director, Daud Abdullah, told CNN he would investigate, and Nasim Ahmed of the Monitor told NPR the outlet is “reviewing our submissions process and strengthening our editorial safeguards.” The outlets were unwitting conduits, not collaborators.

The Production Line

The Iranian influence campaign’s operators prompted ChatGPT in Persian, generated content in Persian and English, and routed everything through VPNs to hide their location. Then they ran the model like a newsroom staff that never slept, a semi-autonomous loop of drafting, tailoring, pitching, and reporting, the kind of agentic workflow we break down in our AI agents vs. agentic AI. First, it drafted and refined the articles: operators asked ChatGPT to polish English drafts and tailor them against specific outlets’ submission criteria, effectively reverse-engineering each publication’s editorial bar. Second, it wrote the pitch emails to editors, automating the least glamorous part of freelance journalism. Third, after publication, ChatGPT generated batches of social media comments to make the articles look widely read and argued-over.

The comments followed a script: the US as aggressor, Israel as an unreliable ally, Iran as defiant victim. Alongside came 20+ batches of hostile Persian replies aimed at Iran International, the UK-based satellite network, plus batches praising the operation’s own articles: a homemade audience, applauding on cue.

But the Iranian influence campaign’s commenting operation flopped. The replies OpenAI could find in the wild drew low engagement, with many of the likes coming from accounts linked to the operation itself. Worse for whoever was paying, the operators’ internal reports measured success with what OpenAI called a deceptive calculation: they counted views on the posts they had replied to, not views of their own replies. That’s like claiming credit for the crowd at someone else’s rally because you heckled from the back row. On OpenAI’s six-point Breakout Scale, the article-planting workstream rated a Category 4. The commenting managed a Category 2.

Dark Clark: The Russian Influence Operation’s Fake Think Tank

The Russian influence operation was stranger and, by OpenAI’s own rating, more dangerous. The company named it “Dark Clark” after a fabricated persona, “Mia Clark,” through whom the operators appeared to control a self-styled research platform, the Social Research Center, in Latin America. Its site claimed a focus on the Indian diaspora in the region; OpenAI found 60+ mostly-original articles on it. The catch: the on-the-ground employees producing that research had no idea they worked for a Russian group.

A cluster of ChatGPT accounts originating in Russia drove the campaign. Most operators prompted in Russian, one in Spanish from someone apparently located in Russia, all connecting through VPNs since OpenAI doesn’t allow access from Russia. ChatGPT wasn’t mainly writing propaganda here. OpenAI’s report says that in the majority of cases, the operators used the model to draft and update internal reports to an unknown superior, status updates, hiring decisions, pay scales, staffing plans. The AI was the operation’s secretary, not its author. That detail alone should retire the lazy narrative that “AI ran the whole thing.” Humans ran it. The AI took minutes.

The operation ran three workstreams: denigrating Ukraine and undermining its armed-forces recruitment, interfering in domestic politics (especially Bolivia and Argentina), and managing the think tank. The operators’ reports claimed fakes from a false story about Argentina’s president buying jeweled dog collars to fabricated audio attributed to Ukraine’s consul in Ecuador, some drawing real press coverage and even a Polish MEP’s response.

But OpenAI urges skepticism about the operators’ own scorecard: some reports show them asking ChatGPT to identify incidents they could take credit for without any involvement, including Argentina’s long-standing Falklands position. OpenAI’s read: the operators appeared to be running “an influence operation targeting their own employers, rather than any external audience.” Even so, Dark Clark rated Category 5, the first Category 5 operation OpenAI has disrupted since it began its reporting, for prompting public comment by politicians in several countries. The company also called it the most complex front-identity attempt it has disrupted in two and a half years of OpenAI influence operations reporting.

The Attribution Problem Nobody Can Solve

Here is the critical nuance, and OpenAI stated it plainly: the company could not identify who was actually running the Iranian influence campaign, or whether it had any link to the Iranian government. “The overall activity appeared consistent with a commercial actor running a for-hire influence campaign,” the report says, though OpenAI added that it was unable to identify the particular actor involved.

If this Iranian influence campaign was bought rather than built, propaganda as a product, attribution gets much harder. A government wouldn’t need to run it, only fund it. And every response, from sanctions to takedowns, gets more complicated when you can’t name the customer, one reason institutional fights over AI control keep escalating, the Pentagon’s public split with Anthropic over Claude’s usage limits is one visible example. Worth remembering the next time someone confidently tells you exactly which capital ordered what.

Why a Real Byline Beats a Thousand Bots

Of the 30 covert influence operations in two and a half years of OpenAI influence operations reporting, the Iranian influence campaign is the clearest demonstration of the company’s own finding: operations that land their content in real media outlets, rather than relying on fake social media distribution, tend to have the highest potential reach and impact. A planted article borrows the outlet’s credibility; readers never see the Persian prompts behind the byline.

What struck me was Jessica Brandt’s word choice. Brandt, who ran the US intelligence community’s now-closed Foreign Malign Influence Center, called it “remarkable” that the operators placed content in genuine outlets, proof AI lets adversaries launder messaging through authentic channels while hiding their hand.

Darren Linvill, who studies influence operations at Clemson’s Media Forensics Hub, told NPR that AI “helped them create language that looked and felt and read authentic to the point where it was being published by legitimate outlets.”

Exposure has ended these fronts before: “Alice Donovan,” a fake journalist persona tied to Russian military intelligence about a decade ago, and “PeaceData,” the fake outlet Meta exposed six years ago, both stopped once exposed. AI made the work faster and harder to spot, but secrecy is still the whole game. The same day, Anthropic gathered its own anti-deception rules into one policy section; the labs are comparing notes in a wider scramble to govern what their systems get used for, a tension we’ve tracked through OpenAI’s misalignment saga.

Reading the News After Bogus Bylines

So what do you actually do with this? You don’t need forensic software, you need habits. The sharpest line I read came from the British outlet ResultSense, which noted dryly that a polished pitch from an unfamiliar freelancer now proves little, checking a writer’s identity matters more than checking their prose.

Start with the byline, not the headline, search the author’s name, because real journalists leave trails. The Bogus Bylines personas had social accounts, but one transparency panel placed the “American freelance writer” in Iran, and that single mismatch unraveled the front. Then read the outlet skeptically: the operation targeted small-to-medium publications big enough to have audiences but small enough to vet thinly. A professional-looking site is not proof of editorial standards.

Then look at the engagement, not the content. Detection tools help at the margins, watermarking schemes like Google’s SynthID flag content from Google’s own AI tools, not ChatGPT output, and watermarks can be stripped anyway, so treat them as one signal among several. AI comment campaigns perform enthusiasm until you notice the engagement is low and the likes come from the same cluster of accounts. Manufactured consensus has a hollow ring if you listen for it. ChatGPT disinformation reads clean in the prose, catch it in the byline instead.

And notice the emotional tuning. AI-generated propaganda, at least in this case, announced itself not in the writing but in the metadata, yet it was engineered to push specific buttons all the same. If a piece makes you furious or vindicated within two paragraphs, slow down and verify before you share it.

One more thing worth saying plainly about this Iranian influence campaign, because the coverage keeps getting it wrong: the AI didn’t go rogue here. No model woke up and decided to plant stories. Human operators wrote the prompts, managed the personas, pitched the editors, and faked the metrics, ChatGPT was the typewriter, not the author. That’s less cinematic than “AI propaganda machine,” but it’s the truth, and it matters: the defenses that work are human too. Vetting, verification, and that unfashionable habit of checking before sharing.

FAQ

References

1. OpenAI, Disrupting AI-enabled “false front” operations, October 8, 2026 (primary source)
2. CNN, Iranian operatives used AI to plant fake stories in multiple US media outlets, October 8, 2026
3. NPR (via WMOT), OpenAI caught Russians and Iranians using ChatGPT for influence campaigns, October 8, 2026
4. NDTV, OpenAI Says It Shut Down Russian, Iranian Propaganda Networks Using ChatGPT, October 2026
5. Unite.AI, OpenAI Bans Two Covert Influence Operations Using False Fronts, October 8, 2026
6. Channel News Asia, Russian fake think-tank with real employees exposed in ‘most complex’ influence operation attempt, OpenAI says, October 8, 2026
7. Cellcog, OpenAI’s False-Front Report: Its First Category 5 Takedown, October 2026
8. ResultSense, OpenAI disrupts Russian and Iranian false-front operations, October 9, 2026
9. Melissa Brodsky (Substack), Iran Didn’t Need to Buy an American Newspaper. It Just Needed a Byline., October 2026
10. The Cyber Express, Covert Influence Operations: How Russia, Iran Used ChatGPT, October 2026
11. AI Affairs, OpenAI bans Russian and Iranian influence operations using false fronts, October 2026