OpenAI Disrupts Iranian Influence Nexus Using ChatGPT

OpenAI has identified and banned five ChatGPT accounts used by Iranian-linked influence operations to generate propaganda and deceptive content. This discovery is significant because it reveals a previously unreported operational link between two separate influence networks, STORM-2035 and IUVM, at the operator level.

Operational Links Between STORM-2035 and IUVM

OpenAI discovered that a single ChatGPT account was used to generate content for both STORM-2035 and IUVM. While the scale of the activity was small, this overlap suggests a shared operational relationship between these two entities that had not been previously reported in public research.

Cross-Platform and Multilingual Content Generation

The identified accounts were used to produce content in multiple languages for various websites and social media platforms:

  • STORM-2035 Activity: Four of the five accounts generated content for entities connected to STORM-2035. This included long-form articles for al-sarira[.]com and Spanish-language articles for lalinearoja[.]net.

  • Social Media Integration: Two accounts generated tweets for the al-Sarira X account and two other X accounts.

  • IUVM Integration: One account generated French-language texts for critiquepolitique[.]com (linked to STORM-2035) and English-language articles for iuvmpress[.]co (linked to IUVM).

Evasion Tactics and Content Themes

Operators utilized specific techniques to avoid detection and maintain the narrative consistency of their campaigns:

  • Semantic Rewriting: In one instance, operators generated articles using OpenAI models and then rephrased them before publication. Semantic similarity analysis confirmed the published versions were highly likely rewrites of the AI-generated text, suggesting a deliberate attempt to evade detection.

  • Narrative Focus: The content was primarily pro-Palestinian, pro-Hamas, and pro-Iran, while opposing Israel and the United States. During the collapse of Bashar al-Assad's regime in Syria, the operation produced content praising Assad and denying his unpopularity.

  • Operator Profiles: Some banned accounts were also used to create materials for teaching English and Spanish as a foreign language. OpenAI notes that this aligns with previous public attributions of Iranian threat activity to individuals with backgrounds in language instruction.

Impact Assessment

Despite the use of AI, the operation failed to gain significant traction. Using the Brookings Institution's Breakout Scale, OpenAI assesses this activity as being at the low end of Category 2. This rating indicates that while the activity occurred across multiple platforms, there was no evidence that real people widely shared or engaged with the content.

As of January 16, 2025, the al-Sarira X account had only 157 followers and typical tweets received only single-digit engagement numbers.

Sources