The network of Iranian influence: Activity spanning multiple platforms

OpenAI Removes Iranian Accounts That Used Its AI for Influence Campaigns

Meta description: OpenAI banned Iran-linked accounts that used ChatGPT to write influence content tied to IUVM and STORM-2035 operations.

Tags: OpenAI, AI misuse, influence operations, STORM-2035, content moderation

Featured image: A friendly abstract illustration of a shield deflecting scattered glowing text fragments, in soft blues and oranges.

OpenAI has shut down several accounts connected to Iran after discovering they were using its AI tools to mass-produce articles and social media posts. The company said the content was linked to two known influence networks: IUVM, a long-running Iranian media operation, and STORM-2035, a group previously flagged for pushing divisive political content aimed at US audiences. This article explains what happened, why these campaigns matter, and what it tells us about how generative AI is being used — and policed — in online influence work.

What OpenAI found

According to OpenAI, the banned accounts were using its models to generate written content at scale. That included articles and social media posts designed to look like they came from legitimate commentators or news outlets. The operation connected to IUVM, which stands for International Union of Virtual Media, has been active for years distributing Iranian state messaging, while STORM-2035 gained attention for posting AI-assisted political content targeting voters in the United States.

OpenAI did not find evidence that the AI-generated material achieved significant real-world reach or engagement. In other words, the campaigns were producing a lot of content, but there was little sign that many people were actually reading or sharing it. That detail matters, because it suggests the biggest challenge right now is the ease of mass production rather than proven effectiveness.

Why this matters

This ban is a good example of how generative AI lowers the cost of running an influence campaign. Writing dozens of articles and hundreds of social posts by hand takes time and staff. With large language models, a small team can produce the same volume in a fraction of the time. The concern for platforms and researchers is that this makes it cheaper for state-linked groups to flood the information environment with misleading or divisive material.

At the same time, the fact that OpenAI caught and removed these accounts shows the other side of the story: AI companies are actively monitoring how their tools are used and cutting off access when they detect coordinated misuse. OpenAI has said it shares its findings with researchers, platforms, and governments so the wider community can track these networks across other services.

The limits of detection

It is worth being honest about what enforcement like this can and cannot do. Banning accounts stops a specific operation on one platform, but determined actors can create new accounts, switch providers, or use open-source models that have no central gatekeeper. Detection also relies on spotting patterns, which works best against campaigns that are large and repetitive — exactly the kind OpenAI described.

There is also the question of attribution. Linking accounts to a state is difficult, and companies like OpenAI base these claims on behavioral signals, technical clues, and outside threat intelligence. Their conclusions are taken seriously, but readers should remember that attribution in this field is rarely simple or certain.

What platforms and users can do

For the companies building these tools, the takeaway is that usage monitoring needs to be constant, not occasional. OpenAI's approach — analyzing how accounts use its models, terminating those engaged in coordinated deception, and publishing reports — has become a template other AI providers are starting to follow.

For everyday users, the lesson is more practical. Content that appears to be written news or commentary may be machine-generated at scale by groups with an agenda. A few habits help: check whether a news site has a real identity and history, be skeptical of posts that push strong emotions during elections or crises, and cross-check surprising claims against trusted sources before sharing.

The bigger picture

OpenAI's action against these Iran-linked accounts fits into a wider pattern. Since 2024, AI companies and social platforms have repeatedly disrupted influence operations from several countries that used generative AI to draft, translate, or polish propaganda. So far, most publicly reported campaigns have struggled to build genuine audiences. But the tools keep improving, and the gap between producing content and winning attention may shrink over time.

The STORM-2035 and IUVM takedowns are best read as an early round in an ongoing contest. Mass AI-generated influence content is now easy to make and relatively easy to catch. The harder problem — one that will require cooperation between AI developers, social platforms, researchers, and informed readers — is keeping that balance as both the campaigns and the defenses evolve.

Who should follow this story

If you work in trust and safety, journalism, or platform policy, this report is directly relevant and worth reading in full. If you are a regular news consumer, the practical point is simpler: treat sudden waves of polarizing online content with healthy skepticism, verify before you share, and expect AI companies to keep announcing takedowns like this one as their monitoring tools mature.