Why Generative AI Still Struggles to See Africa Beyond the Stereotypes

A Nigerian designer has proposed a five-part method for examining who generative AI represents, how accurately it portrays different cultures, and which users must work harder to be recognized.

When user experience designer Joy Bachashe Samuel asked an artificial-intelligence model in 2024 to generate a young Nigerian boy via a text-to-speech prompt, the result was not recognizably Nigerian. Instead, the system produced what she considered a generic collection of features commonly associated with an “African child”. A second experiment exposed a different pattern. Samuel requested images of a renowned surgeon. All four results showed white surgeons. “I asked for a young Nigerian boy, but what I received looked like a generic idea of an African child rather than a Nigerian boy,” Samuel said. “When I prompted for a renowned surgeon, all four images generated by the model showed white surgeons.” Her experience reflects concerns already documented by researchers studying text-to-image systems. Research by Federico Bianchi and colleagues found that ordinary prompts involving occupations, personal characteristics, and social roles could reproduce or intensify racial and gender stereotypes even when the prompts did not mention identity. Another study, Stable Bias, examined three widely used text-to-image systems and found that they underrepresented marginalized identities to differing degrees. Its researchers proposed comparing outputs across professions, ethnic markers, and gender indicators to identify measurable representation patterns. Samuel’s contribution enters this wider discussion from an African product-design and user-experience perspective.

 

Published in January 2025, Samuel’s open-access white paper, Representing Africa Accurately: A Framework for Evaluating Visibility, Cultural Fidelity and Equality in Generative AI, proposes the concept of Representational Accuracy. The aim is to confirm whether an AI system depicts people and cultures with appropriate visibility, equal social context, cultural credibility, comparable quality, and reasonable accessibility.

Samuel argues that the problem is not limited to whether an AI model can generate a Black person. The setting, cultural details, profession, visual quality, and social role attached to that person must also be examined. This is especially relevant when systems treat Africa as culturally uniform. Images may combine unrelated clothing, architecture, and geographic features, or present African subjects in overly traditional or rural settings even when users request contemporary people and environments.

 

A central idea in this paper is the Identity Tax. It describes the additional work some users perform before an AI system produces a culturally credible result. That work may include writing longer prompts, adding racial or cultural keywords, correcting stereotypes, and regenerating an output several times.

“It exposed a design failure because getting one quick output took much longer than expected,” Samuel said. “I had to write increasingly detailed prompts and add words such as ‘African’, ‘Saharan’ and ‘dark-skinned’ before the results moved closer to what I needed.”

The paper also introduces Phenotypic Defaulting, a term describing how identity-neutral prompts can repeatedly resolve towards particular dominant physical characteristics. The framework recommends a six-stage audit process which first defines the system, users, markets, languages, and communities being assessed. They then create neutral, professional, positive, negative, cultural, and matched prompts.

This focus on structured evaluation aligns with the broader movement to test and document generative-AI risks. The United States National Institute of Standards and Technology, for example, has published a generative-AI companion to its AI Risk Management Framework to help organizations identify and manage risks associated with generative systems.

 

Samuel’s framework is currently a proposed method, not an internationally recognized standard. Its usefulness will therefore depend on how researchers and product teams test it in practice. Important questions remain, including how to measure cultural accuracy, who should evaluate it, and how organizations can avoid imposing rigid identity categories on generated people.

Samuel wants technology companies to incorporate representational testing before products reach users.

“I want companies to carry out culturally informed testing and include African users in the evaluation process,” she said. “Bias should be measured before a product is released, and designers should recognize that representation is part of their responsibility.”

Whether the proposed framework becomes widely adopted will depend on independent testing. For now, it adds an Africa-centered product-design perspective to a growing body of evidence showing that generative AI does not always treat a neutral prompt neutrally.

The white paper is available through https://doi.org/10.5281/zenodo.23185753