Covert influence operation: Ghana election activity
OpenAI banned accounts using AI to generate articles, posts, and fake engagement around Ghana’s 2024 presidential election.
This case study was originally published in OpenAI’s February 2025(opens in a new window) report.
Actor
We banned a cluster of ChatGPT accounts that were generating short comments and long-form articles about the Ghanaian presidential election in English. The content supported one candidate, Vice-President Mahumudu Bawumia, and criticized his opponent, former President John Mahama. The activity, including the website Empoweringghana[.]com at the center of the operation, was linked to a commercial entity with offices in the United Arab Emirates and Ghana, named DigitSol. We banned the accounts before the election was held.
Behavior
This operation used ChatGPT to generate English-language content that was then posted across the internet.
At the core of the operation was a website that claimed to represent a youth initiative called “Empowering Ghana.” The operation used ChatGPT to generate articles for this website. Although the site’s content focused exclusively on Ghana, its “contact” section provided a street address in Australia, an apparently invalid phone number, and a link to a WordPress plugin provider.
The website, in turn, linked out to branded “Empowering Ghana” accounts on Facebook, Instagram, X, YouTube, TikTok, and WhatsApp. The operation used ChatGPT to generate many of the posts on these social media assets.
The operation then used our models to create short comments that were posted in reply to the branded “Empowering Ghana” social media posts. For example, the operation would generate 30 short comments praising one of Vice-President Bawumia’s policies. All 30 comments would then be posted by a single social media account on Instagram or X. Some of the social media accounts doing this commenting had no profile pictures; others had pictures copied from across the internet.

Comments on Instagram (left) and posts on X (right) generated by this operation using ChatGPT and posted in reply to posts by branded “Empowering Ghana” assets.
The website and social media accounts appear to have published their last posts on December 8, 2024, the day after the election.
Completions
The content generated by this operation ranged in length from four-word comments to lengthy news articles in English. The majority of the content praised or supported Vice-President Bawumia or criticized or opposed former President Mahama.
Impact
Assessing the impact of this operation requires particular care, because there are some indications that the operators resorted to fake engagement to exaggerate the impact of their activity. For example, many of the posts on the “Empowering Ghana” Instagram account received 30-32 comments, all of them pro-Bawumia. However, as noted above, the operators regularly used our models to generate sets of 30 comments, and on some occasions we identified all 30 among the replies to a single “Empowering Ghana” post. As such, it is likely that the authentic engagement on most of those posts was in the realm of 0-2 comments, rather than 30-32.
Similarly, the Empowering Ghana Instagram account apparently had around 900,000 followers. However, Instagram’s own transparency settings indicated that 54% of those followers, about half a million accounts, also followed one other specific Instagram account that focused on Australian legislation. The likelihood that some 500,000 authentic Instagram users with a focus on Ghanaian politics organically followed a single other account with a focus on Australian legislation appears remote. While open-source investigation cannot determine this with certainty, the Instagram following appears more likely to have been artificially inflated.
By contrast, the Empowering Ghana X account had 1,535 followers, and the YouTube channel just over 1,500 subscribers, as of January 16, 2025. Typical tweets from the Empowering Ghana account received engagements in the single or low double digits, typical videos in the double or low treble digits. Our investigation did not identify occasions on which Empowering Ghana content was amplified by mainstream media.
Using the Breakout Scale(opens in a new window) to assess the impact of IO, which rates them on a scale of 1 (lowest) to 6 (highest), we would assess this as being in Category 2 (activity on multiple platforms, but no evidence that real people picked up or widely shared their content).
Domain Associated With This Operation
Empoweringghana[.]com.