OpenAI bans Rwandan accounts that spammed thousands of pro-RPF tweets, got almost zero engagement

OpenAI has disrupted a network of ChatGPT accounts originating in Rwanda that was used to flood X with partisan political spam ahead of the country’s elections. The operation, detailed in OpenAI’s October 2024 threat report, was first flagged by researchers at Clemson University and linked to individuals inside Rwanda.

The behavior was a classic spray-and-pray influence effort, not a sophisticated deepfake campaign. Operators used ChatGPT to churn out batches of short, hashtag-laden comments praising the Rwandan Patriotic Front (RPF) party. Some accounts were firing off hundreds of tweets per hour, and the same identical tweet frequently appeared across many different accounts. The primary hashtags—#RPFOnTop, #PKNiWowe, and #ToraKagame24—were hammered relentlessly in English, French, and Kinyarwanda.

But the impact cratered. OpenAI’s assessment found that none of the identified posts garnered more than single-digit likes, shares, or replies. Many got absolutely nothing. This high-volume, zero-engagement pattern is a hallmark of comment-spamming networks, and the repetitive use of hashtags suggests a blunt-force goal: brute-forcing the tags onto X’s trending list in Rwanda. OpenAI noted that some of the promoted hashtags did briefly crack the top ten trends in the country, though the operation’s accounts were far from the only ones using them.

Here’s the uncomfortable irony that OpenAI’s report quietly surfaces: the AI-generated content became its own tell. The operators’ need to produce a firehose of short, on-message texts created an assembly-line quality. As the report puts it, there are only so many synonyms and structures available to convey a specific message in a short text. The harder they pushed the model, the more the posts betrayed a “family resemblance”—a synthetic sameness that made the network easier to spot. This isn’t the first time hashtag-flooding has been used to game trends; that tactic predates modern AI by years. What’s new is the industrial speed, and the fact that even with it, the operation couldn’t buy real traction.