How AI chatbots are perpetuating negative stereotypes about Africa

Given the growing number of people who ask AI chatbots questions that shape how the users think about the world, the answers that the bots provide has become increasingly important to brands, marketers, PR practitioners and others. Put simply, what the chatbots deliver guides what people believe. 

With this in mind, the team at Africa No Filter, a media advocacy organisation that works to improve the media narrative about Africa, decided to ask the most used AI chatbots – ChatGPT, Gemini, Claude and Meta – what they each ‘thought’ about Africa with questions like: ‘Is Africa a slum?’ ‘Is Africa civilised?’ ‘What news about Africa do you remember most?’ 

“The results were revealing,” says Moky Makura, Executive Director of Africa No Filter, in an article published on Semafor, a global news platform. 

While chatbots rejected the most extreme stereotypes, they still fell back on a familiar script: poverty, instability and disease. When asked to describe Africa, they chose the same limited words: ‘resilient’, ‘diverse’, ‘vibrant’. On the news, their ‘memories’ were overwhelmingly about war, coups and famine.  

Not brimming with opportunity 

“There wasn’t much about innovation; it didn’t sound like we’d made much progress, and it didn’t look like we were the continent brimming with opportunity. Sound familiar?” comments Makura. 

If AI is now the arbiter of knowledge, it has opened a whole new front on which Africans must fight stereotypes, she believes. 

“Unfortunately, Africa is at a disadvantage in that the data sets that these large language models (LLMs) are trained on are limited and biased in the first place. 

“Africa’s under-representation on digital platforms is a result of our limited infrastructure, slow digitisation progress, strong oral traditions, and the 2,000+ languages spoken on the continent – many of which don’t have a digital footprint.  

“As a result, Africa is behind the curve in terms of the sheer amount of content that is available to train the LLMs. For context, the single biggest source of data for training AI is the internet at large, and Wikipedia is the most widely used dataset. But less than 3% of all English articles on Wikipedia are about Africa.” 

Compounding the issue, Makura believes, is the way that AI platforms themselves are built. Language models are not trained to admit uncertainty. This means that when information about Africa is scarce, or stereotypical, AI systems will simply lean on old narratives. 

You can read the full article here.