OpenAI Disrupts Operation A2Z Multilingual Influence Activity

OpenAI has disrupted "Operation A2Z," a multilingual influence operation that used the OpenAI API to automate the creation and management of fake social media personas. The operation leveraged AI to generate bios, analyze posts, draft multilingual replies, and produce stylized images to influence political discourse on X and Facebook.

AI-Driven Persona Management and Content Generation

Operation A2Z utilized AI to automate multiple stages of the influence "killchain," allowing a small number of operators to manage approximately 150 identified accounts across X and Facebook. The AI was used for the following specific tasks:

  • Persona Development: Generating bios for fake accounts to establish credibility.
  • Content Creation: Drafting short comments and replies in multiple languages.
  • ** uma Proofreading:** Refining generated text for better flow and authenticity.
  • Engagement Analysis: Analyzing existing posts and comments to tailor responses.
  • Visual Content: Creating stylized images, often resembling 1930s posters, designed to attract attention rather than deceive.

By automating these processes, the operation avoided the need for a large team of human "trolls," reducing operational costs and theing risk of leaks.

Targeted Themes and Political Narratives

While the operation focused on a wide range of topics, suggesting it may have been a commercially run entity, several prominent political themes emerged:

Azerbaijan Support

The most frequent activity involved praising the government of Azerbaijan. This included promoting Azerbaijan's role as an energy supplier to Europe, defending its human rights record, and highlighting its role in hosting the UN Climate Change Conference and promoting unity in Central Asia.

Regional and Global Stability

  • United Arab Emirates: Some accounts portrayed the UAE as a stable investment destination, often contrasting this with claims of political instability in the United States.
  • European and US Politics: The operation deployed varying ideologies depending on the region. For example, some X accounts posed as liberals to criticize former US President Trump, while French-language Facebook accounts criticized President Macron and supported the National Rally (RN) party.
  • International Intervention: Some content argued that Western countries should focus on domestic issues rather than intervening in international conflicts, specifically citing Russia's invasion of Ukraine.

General Interest Topics

Beyond politics, the operation generated content regarding financial markets, cryptocurrencies, ecological issues, the evolution of AI, and general technology.

Impact Assessment and Vulnerability

OpenAI assesses the impact of Operation A2Z using the Breakout Scale, rating it at the top end of Category Two (on a scale of 1 to 6), with a risk of moving into Category Three. This rating is based on the operation's ability to post across multiple platforms and engage real users in conversations.

Engagement Metrics

Despite the high volume of activity, the operation's reach was limited:

  • Followers: The largest X account identified had 222 followers; most had between 15 and 25.
  • Engagement: Typical posts on X received single-digit engagements. The highest number of reactions to a Facebook post was 36, while most ranged from 0 to 5.

The "AI Vulnerability" Paradox

OpenAI notes that while AI increases the capability of influence operators, it also creates a critical vulnerability. Because Operation A2Z relied on AI at almost every stage of its operation, banning the API access effectively broke multiple links in the killchain simultaneously.

Following the disruption in early June, the identified accounts stopped posting during key election periods in the EU, UK, and France. Although some resumed posting in late August using other AI models, the original scale and efficiency were significantly diminished.

Sources