OpenAI Operation No Bell Case Study

Operation No Bell: AI-Driven Influence Operation Targeting Africa

OpenAI has banned a ChatGPT account linked to a Russian-based influence operation, dubbed "Operation No Bell," which used generative AI to produce long-form articles and social media content criticizing the US and its allies. The operation focused on sub-Saharan African geopolitics, utilizing fake academic personas and repurposed social media pages to plant disinformation in mainstream news outlets.

Actor and Origin

The operation likely originated in Russia, with investigations triggered by a lead from Meta. The name "No Bell" stems from a specific instance where the ChatGPT user explicitly prompted the model to advocate for the President of Angola to win the Nobel Peace Prize, stating that the goal was to antagonize the team of US President Donald Trump.

Content Generation and Distribution Tactics

Operation No Bell employed several technical and social engineering tactics to bypass detection and increase the perceived legitimacy of its content:

Use of Fake Personas

Many long-form articles were published under the byline of "Dr. Manuel Godsin," who was claimed to be a PhD holder from the University of Bergen and an employee of the "International Centre for Political and Strategic Studies." OpenAI's investigation found no record of such a person in the Norwegian National Research Information Repository (NVA) or the Bergen University Library. Further research revealed that the photo used for Godsin was actually that of a St. Petersburg law student from the 2010s.

Evasion of AI Detection

To conceal the AI-generated nature of the text, the operator used specific prompting techniques:

  • Stylistic Constraints: The user prompted the model to write in the style of a human journalist.
  • Formatting Removal: The user specifically requested the removal of em-dashes, a common marker of AI-generated text. In one instance, a published article in Ghana contained garbled text where em-dashes had been manually removed from the original AI generation.

Cross-Platform Coordination

The operation utilized multiple AI tools and platforms:

  • Facebook Pages: Content was posted to Facebook Pages posing as news outlets in South Africa, Ghana, Kenya, and Angola. Some pages were acquired and repurposed; for example, "Kenya Watchtower" was previously named "Farmtown5."
  • Multi-LLM Usage: The user experimented with different models, requesting image prompts for Gemini and asking ChatGPT to modify images that appeared to have been generated by Gemini for use as Facebook banners.

Thematic Focus and Narrative Goals

The generated content primarily praised Russia and criticized Ukraine, the United Kingdom, and the United States. Key narratives included:

  • Geopolitical Criticism: Articles praised Russia's presence in the Central African Republic and accused Western leaders of targeting South Africa with disinformation.
  • Localized Issues: The operation targeted specific entities, such as accusing the German arms manufacturer Rheinmetall of circumventing export controls via a South African subsidiary and accusing the British NGO Crisis Action of fomenting protests in South Africa.
  • Personalized Attacks: Some content focused specifically on Ukrainian President Volodymyr Zelenskiy or US President Donald Trump.

Impact Assessment

While the operation had limited reach on social media—with one page reaching only 3,000 followers before being banned by Meta—it was more successful in planting articles on mainstream news websites.

Using the Brookings Institution's Breakout Scale for influence operations (which ranges from 1 to 6), OpenAI assesses Operation No Bell as being toward the low end of Category 4, indicating that the operation successfully broke out into mainstream media.

Sources