OpenAI announced the disruption of an influence operation it named "A2Z," which utilized its AI models to create multilingual content for social media. The operation involved approximately 150 accounts across X and Facebook, with more suspected. These accounts generated comments praising Azerbaijan, defending its human rights record, and addressing a range of other topics, suggesting a potentially commercially run operation. This case study was originally published in OpenAI's October 2024 report.
Key Points
- OpenAI banned "A2Z" accounts that used its API to generate multilingual influence content.
- The operation focused on elections, Ukraine, and politics, with a prominent theme of praising Azerbaijan.
- AI was leveraged to manage fake social media personas, generate bios, analyze posts, draft replies in multiple languages, and proofread content.
- The content, primarily short comments and stylized images, was posted on X and Facebook.
- The operation involved around 150 identified accounts across X and Facebook.
- Disruption occurred in early June, leading to a halt in posting during election periods in the EU, UK, and France.
- Some accounts resumed posting in late August, with content showing indications of AI generation, though not from OpenAI models.
Context
According to OpenAI, the "A2Z" operation used AI to enable social media commenting, similar to the "Bad Grammar" operation it exposed in May. The use of AI allowed the operation to manage numerous social media accounts simultaneously and engage in short conversations with real people. OpenAI also reported on "Uncle Spam" in its June 2025 report, an operation linked to China that generated polarized US political content, and "Helgoland Bite," also from its June 2025 report, which involved Russian-origin accounts creating German-language political content. Additionally, OpenAI's October 2024 report detailed "STORM-2035," an Iran-origin operation generating US and UK election content.
Why It Matters
This disruption demonstrates how AI can be used to scale influence operations, enabling a smaller team to manage a large number of fake social media personas and generate diverse content. It also highlights the vulnerability of such operations to AI providers' detection and disruption efforts, as reliance on AI at multiple stages of content creation can lead to a cascading takedown.
What To Do
- Note that OpenAI's October 2024 report contains further details on the "A2Z" and "STORM-2035" operations.
- Watch for additional reports from OpenAI on malicious uses of AI, such as those detailed in the June 2025 report regarding "Uncle Spam" and "Helgoland Bite."
- Compare the tactics used by "A2Z" (multilingual, praise for specific countries) with those of other operations like "Uncle Spam" (US polarization) and "Helgoland Bite" (German-language political content).