Bias and Fairness in AI Systems
AI systems learn from data — and data comes from the real world. When the real world contains unfairness, AI can learn that unfairness and repeat it at scale. Understanding bias in AI is not just a tech topic; it is a social justice topic.
What Is Bias in AI?
Bias in AI means the system produces results that are systematically unfair to certain groups of people. This is not usually intentional — it happens because the data used to train the AI, or the way the problem was set up, reflected existing inequalities.
A classic example: a facial recognition system trained mostly on photos of light-skinned people will be less accurate at recognising dark-skinned faces. This is not because someone decided to build a racist system. It is because the training dataset was not representative — and nobody caught the problem before deployment.
1. Data bias: The training data does not represent all groups equally.
2. Algorithmic bias: The way the AI optimises its goal advantages some groups over others.
3. Deployment bias: The AI is used in a context very different from where it was designed, causing harm.
Why This Matters in Kenya
Most powerful AI systems in the world were built in the United States or China, trained mostly on English-language data, and tested on users in high-income countries. When these systems are deployed in Kenya, the bias problems can be severe.
How to Recognise Biased AI
- Who was the AI trained on? Ask whether the training data included people like you — your language, your economic context, your geography.
- Who tested it? Was the AI evaluated on a diverse group before launch, or only on a narrow group?
- Who is harmed by errors? If errors consistently hurt one group more than another, that is a bias signal — even if the overall accuracy looks good.
- Does it perform differently for different groups? Ask for disaggregated accuracy data (accuracy broken down by gender, region, language, income). If the company cannot or will not share this, that is a red flag.
Fairness — More Complicated Than It Sounds
Here is something that surprises many people: there is no single mathematical definition of "fair." Researchers have identified over 20 different mathematical definitions of fairness, and they are often contradictory. You cannot make an AI satisfy all of them at once.
This means fairness in AI is ultimately a social and political question, not just a technical one. Who gets to decide which definition of fairness to use? Who is represented in that decision? In Kenya, this means citizens, civil society, and government all have a legitimate role in demanding accountability from AI systems — not just engineers.
Bias Is Not Permanent
The good news: bias in AI can be identified and reduced. It requires more diverse training data, more careful evaluation across groups, diverse teams of developers, and strong accountability structures. This is why the AI literacy you are building right now matters — the more Kenyans understand these issues, the more pressure there will be to build AI that works fairly for everyone.