Operation “No Bell”: Coordinated criticism of the US and allies

This case study was originally published in OpenAI’s February 2026⁠(opens in a new window) report on disrupting malicious uses of AI.

Actor

We banned a ChatGPT account that was generating long-form articles and social media content about geopolitics in Africa. We began investigating this operation following a lead from Meta. The account likely originated in Russia. One of the long-form articles advocated that the president of Angola should win the Nobel Peace Prize; the ChatGPT user stated explicitly that this was intended to antagonize the team of US President Donald Trump. Based on this, we have dubbed it “No Bell”.

Behavior

The ChatGPT account’s main activity was generating social media posts and long-form commentary articles about geopolitics in sub-Saharan Africa. The user mainly prompted in English, but sometimes input Russian-language instructions that they attributed to their manager. These instructions included edits and detailed feedback on some of the articles.

Some of the user’s content was posted on Facebook Pages that posed as news outlets in African countries including South Africa, Ghana, Kenya and Angola. Meta banned these Pages. Under Facebook’s transparency settings, a few of the Pages showed their admin locations, which included Russia and Ukraine. One Page focused on Kenya showed that it had originally been called “Farmtown5” and had then changed its name to “Kenya Watchtower”, suggesting that it had been acquired and repurposed from a supplier unrelated to the operation. We also identified references to previously removed Pages targeting Zambia and Namibia.

Transparency settings for one of the Facebook Pages in the Transparency settings for the Page “Kenya Watchtower”, network, showing the admin locations. The ChatGPT user generated showing the name change. The ChatGPT user generated content that was posted by this Page. Meta banned this Page content that was posted by this Page. Meta banned this Page.

As with operation

“Trolling Stone”, and some of the operations we reported on in October⁠(opens in a new window), this user occasionally prompted the model to generate text without em-dashes, to help conceal the fact that it was AI-generated. On one occasion, we identified one of their long-form articles published by a news website in Ghana, where em-dashes that had featured in the original generation had been removed, leaving a garbled text. They also asked our model to make sure its writing was in the style of a human journalist, likely to reduce the appearance of AI-generated content.