ALGORITHMIC BIAS: AFRICAN STEREOTYPES PORTRAYED BY LLMs
Mfundo Mbambo
This project tests whether popular AI chatbots describe African people differently from Western people. As large language models are increasingly used across Africa to write job summaries, school materials, and news, the way they portray African subjects has real consequences. Yet this kind of geographic bias is rarely measured, and most existing bias research focuses on Western social categories rather than how Africa is represented compared to the rest of the world.
Our approach is simple. We give five widely used AI models matched pairs of prompts that are identical except for the location. For example, "a doctor's day in Berlin" versus "a doctor's day in Nairobi," or "a child's dream in Oslo" versus "a child's dream in Lagos." By comparing the two answers, we can see whether the model treats the African subject differently, and in what ways.
Matched-pair contrastive prompting focused specifically on Africa — While geographic bias in LLMs has been studied, most work focuses on performance gaps (e.g., IrokoBench testing how well models handle African languages) rather than representational bias in how models describe African subjects. Your approach of comparing narrative outputs for identical prompts with only location changed is a clean, intuitive methodology .
Practical downstream framing — Connecting this to real-world use cases (job summaries, school materials, news) in Africa gives it applied significance beyond academic interest .
This paper offers a valuable, Global South–led audit of how widely deployed LLMs narrate African subjects compared to Western ones, with a clear focus on representational harms in everyday scenarios that are directly relevant to AI deployment in African contexts. The paired‑prompt design, four prompt groups, and structured harm rubric make the methodology easy to follow and reproduce, and the inclusion of a non‑Western model strengthens the claim that the observed bias is not just a US‑data artifact.
However, the study’s quantitative robustness is limited by the small prompt set, single‑rater scoring, and absence of inter‑rater reliability or basic statistical analysis; adding a few concrete scoring examples and a short per‑model summary would make the findings more compelling. The writing is generally clear and well structured, with vivid examples, but could be tightened further by explicitly stating the novel contribution and “theory of change” (how this audit can inform procurement, benchmark design, and governance for African institutions) and by presenting the rubric in a more digestible format.
Overall, this is a strong exploratory audit with high relevance for Global South AI safety that would benefit most from scaling up the methodology and sharpening its pathway to concrete policy and evaluation tools.
The finding that models could accurately diagnose the bias they had just produced — confirming it's a hidden default rather than an intentional output — is the single most memorable result in this set and immediately communicable to a non-technical audience. To give the findings real scientific weight, recruit just one additional person to score a portion of the outputs independently, and archive all raw model responses in a public folder so others can verify and build on your work. Your Africa-vs-West framing is complementary to their intra-Africa work, and a collaboration could significantly amplify the impact of what you've started here.
Cite this work
@misc {
title={
(HckPrj) ALGORITHMIC BIAS: AFRICAN STEREOTYPES PORTRAYED BY LLMs
},
author={
Mfundo Mbambo
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
