How Biased Is Your AI? Evaluating LLMs in African Contexts with AfriStereo
Large language models are increasingly being used across Africa, but how do we know whether they reproduce stereotypes about the people and communities they are designed to serve?
Many widely used AI bias benchmarks were developed around Western social and cultural contexts. This leaves an important gap when evaluating models intended for African users.
AfriStereo is a culturally grounded dataset and public leaderboard designed to evaluate stereotype bias in LLMs using stereotypes collected from people in Nigeria, Kenya and Senegal.
In this webinar, we'll explore how AfriStereo was developed, what its Bias Preference Ratio (BPR) tells us about model behaviour, and how researchers, AI builders and product teams can use the leaderboard when evaluating models.
We'll cover
Why culturally grounded AI evaluation matters
How the AfriStereo dataset was developed
Understanding stereotype-anti-stereotype testing and BPR
What happens when stereotypes are explicitly framed as African
A live walkthrough of the AfriStereo Leaderboard
How researchers and organisations can use and contribute to the benchmark
Q&A with the AfriStereo team
Who should attend
This session is designed for:
AI researchers and social researchers · Responsible AI practitioners · AI product managers and AI leads · Startups building AI products for Africa · Students and academics
No specialist knowledge of AI fairness metrics is required.