OpenAI Bans Russia and Iran AI Influence Operations: Fake Journalists, Real Newsrooms

Introduction: When the Byline Is a Bot

On October 8, OpenAI published its latest influence-operations report, and it reads less like a tech disclosure and more like a spy novel. The company banned clusters of accounts traced to Russia and Iran that used ChatGPT to run covert propaganda campaigns — including one that successfully placed AI-written articles in real newspapers under invented journalist personas, and another built around a fake think tank staffed by unwitting real employees.

This isn't a hypothetical about AI misuse anymore. It's a documented case of generative AI being used to launder geopolitical messaging through legitimate media — at scale, for hire, and with bylines you could read for months without knowing they belonged to no one.

Here's what happened, why the mechanics matter more than the geopolitics, and the verification-first AI tools that flip the technology from a liability into a defense.

1. The Iranian Operation: 100 Articles, 7 Fake Bylines

The Iranian-origin campaign, which OpenAI judges to be the work of a commercial for-hire actor, worked like a content farm in reverse. Instead of building fake outlets, it targeted real ones.

Operators sent instructions to ChatGPT in Persian, using the model to refine articles about the US–Iran conflict and then draft email pitches to genuine media outlets around the world. The result: nearly 100 articles published or syndicated across roughly a dozen outlets between July 2025 and October 2026, all under seven fictitious bylines posing as American or Western journalists covering foreign affairs.

One of those pieces — under the invented byline "Ervin Hoskins" — landed in a local Florida newspaper, as The Washington Post reported. Some articles were then promoted by the social media accounts of the outlets themselves, handing the operators distribution they could never have built on their own.

OpenAI rated the media-placement arm of the campaign Category 4 on its five-point influence-operations scale — meaning it achieved meaningful real-world reach. A companion effort, ChatGPT-generated social media comments on the US–Iran conflict in English and Persian, flopped: most drew single- or double-digit engagement, largely from the operation's own accounts.

The lesson is blunt. Audiences have learned to distrust anonymous social media spam. They have not yet learned to distrust a bylined article in a local paper — and that asymmetry is exactly what the operators exploited.

2. The Russian "False Front": A Fake Think Tank With Real Employees

The Russian operation is the one OpenAI calls the most complex front-identity attempt it has disrupted in two and a half years of publishing these reports — its first-ever Category 5 disruption.

The core of the campaign was a front research organization called the Social Research Center, a fake think tank targeting Latin American audiences. Its twist: it hired real, unwitting local employees, giving the outfit a genuine office footprint, real staff producing real-looking output, and a corporate veil that's far harder to pierce than a network of bot accounts.

ChatGPT was used to generate both internal reports and public-facing content for the front, in service of, as OpenAI put it, "laundering geopolitical, conflict-related messaging into their target audiences." Evidence suggests the operation succeeded in prompting public comments by politicians in several countries — the first time OpenAI has documented an influence operation crossing that line since it began these reports.

Category 5 — the second-highest impact rating on the scale — means this wasn't a test balloon. It worked well enough that elected officials repeated its messaging in public.

3. The New Propaganda Playbook

Strip away the geopolitics and both campaigns followed the same three-step playbook, and it's one every AI user should recognize:

• Draft with AI: Use a consumer chatbot to produce fluent articles, reports, pitches, and comments in the target language — cheaply and at volume.
• Borrow legitimacy: Don't build fake outlets. Pitch real ones, staff a real front org, or ride the social accounts of publications that republished your work.
• Invent the human: Seven fake journalist personas, a fake think tank's bylines — the human face is the payload, and the AI is just the typewriter.

Notably, OpenAI has said the campaigns relied on its own models to moderate — the platforms' detection came from usage-pattern analysis and cross-referencing with external researchers, not from the models refusing to write the content. Detection is happening at the account level, not the prompt level. That gap is where media literacy and verification tools have to do the work.

4. How OpenAI Caught Them

According to OpenAI's report, the takedowns combined internal investigation with external verification from disinformation researchers. Accounts were banned before the campaigns were publicly named, and the report deliberately documents infrastructure — personas, front organizations, publication lists — so that journalists and researchers can trace the footprint further.

This is now a routine part of frontier-lab operations. OpenAI has published influence-operations reports for over two years, disrupted two Chinese-linked campaigns in June 2026 alone, and coordinates with peers like Google, whose own threat-intelligence teams track similar abuse of Gemini and Claude. The uncomfortable truth: every capable model gets targeted this way, and the labs find these operations mostly after they've run for months.

5. How to Defend Yourself: AI Tools That Verify Instead of Fool

The same technology that mass-produces believable prose can also be turned against itself. A few categories of AI tools are genuinely useful for spotting laundered content and fake expertise:

Cross-check the source, not just the claim

AI answer engines like Perplexity and Perplexity Pro cite their sources inline, which makes it dramatically harder for a single planted article to masquerade as consensus. If a claim traces back to one byline at one outlet you've never heard of, treat it as unverified.

Interrogate the evidence chain

Research tools like Elicit and Consensus ground answers in indexed academic papers and show you the underlying studies. A fake think tank's "report" won't survive contact with a tool that links every assertion to a real, checkable paper.

Ask the model to attack the article

A practical prompt pattern: paste a suspicious article into ChatGPT or Claude and ask it to identify logical fallacies, unsourced claims, and stylistic markers of AI generation — then ask which specific facts it would need verified. Models are better critics than they are detectors, and a critical pass catches most laundered op-eds.

Verify the byline

The Iranian operation collapsed under a simple question: does this journalist exist? A person with a byline but no history, no other publications, and no professional footprint is the single strongest red flag in both campaigns OpenAI disrupted.

Frequently Asked Questions