How AI Disinformation Works, and How OpenAI Caught Two Operations

On Oct. 8, 2026, OpenAI published a report on two covert influence operations it banned from ChatGPT. One ran a fake research center with an invented analyst. The other posted articles under seven fake journalist names. Both are useful case studies, because they show how AI is used in disinformation today, where it helps operators, and where it does not.

The Two Operations

"Dark Clark" (Russia-based)

• Who and where: ChatGPT accounts operated from Russia, using VPNs, and aimed at Latin America.
• The goal: OpenAI says the aims were to undermine Ukraine and influence politics in Argentina and Bolivia.
• The front: A fake research platform, the Social Research Center, with an invented persona, "Mia Clark." It published more than 60 articles and had 961 LinkedIn followers.
• How AI was used: Mostly for internal reports, with some content such as a fake letter and translations.
• The rest of the playbook: Fake emails to schools in Peru and Ecuador and fake audio, including a clip on Ecuadorian TikTok on April 8, 2026, and a fake about a water shutoff in Bolivia that the utility, EPSAS, debunked.
• Attribution: OpenAI says open-source researchers tied some of the fakes to a group reported as a successor to Wagner. We have not independently verified that link.

"Bogus Bylines" (Iran-based)

• Who and where: Accounts in Iran working in Persian and English, in a pattern OpenAI says is consistent with a commercial, for-hire campaign. The paying actor was not identified.
• The front: Seven fake journalist bylines, with almost 100 articles placed across about a dozen outlets from July 2025 to October 2026, mostly since March 2026.
• How AI was used: To refine articles to each outlet's standards, draft pitches to editors and generate replies. The operators also posted over two dozen batches of comments aimed at Iran International.
• How they measured success: By views on the posts they replied to, which OpenAI says inflated their impact.

How AI-Enabled Disinformation Works

This section is explainer, drawn from patterns in the report. AI is not mainly used to invent a lie from nothing. It lowers the cost of everything around it:

• Volume and polish. A fake article can be written in good English or Persian, matched to an outlet's style, and pitched to editors at scale.
• Personas. A fake analyst or journalist needs a body of work to look credible. AI helps produce it quickly.
• Translation. Operators can work across languages without native speakers.
• Back-office work. In the Russia case, the main use was internal reports, a sign that AI speeds up operations more than it creates persuasive content.

The weakest link is still distribution. Getting real audiences to read and trust the content is the hard part. OpenAI says the operators' own impact claims were unreliable, and that some of their fakes could not be found in open-source searches.

How Platforms Catch Them

OpenAI says it combined its own account data with open-source research. It banned the accounts and shared information with authorities and industry partners. Meta banned a related Instagram account in August. A pattern from earlier cases cited in the report: operations called "Alice Donovan" and "PeaceData" stopped after exposure, which suggests that being named is a real cost to operators.

OpenAI rated the Russia-based operation Category 5 on the Breakout Scale, its framework for how far an operation spread beyond its own channels, and the first Category 5 it has disrupted. The Iranian operation's comments were Category 2 and its articles Category 4. These are OpenAI's own assessments, not independent ratings.

What Readers Can Do

This section is analysis, not reporting. Checks that work against this kind of content are simple: look for a byline's history outside one outlet, search the "think tank" for staff and a physical address, and treat audio or video of dramatic events with caution until a named official or outlet confirms it. Our guide to OpenAI's textGrain watermark explains why a missing watermark proves little, and the AI glossary defines terms like synthetic media.

Why It Matters

• The model makers are publishing takedowns. This is the first Category 5 case OpenAI says it has disrupted.
• AI mainly speeds up the workflow. The Russia-linked case used it largely for reports and translation.
• Distribution is the bottleneck. Operators' impact claims were inflated.
• Ratings are self-assessed. The Breakout Scale scores come from OpenAI itself.