Bias & Fairness

⚖️ What AI bias is

AI bias is when a system treats some groups of people unfairly, for example making more mistakes for women, for darker skin, for older people, or for one region's accent. AI doesn't choose to be unfair: it copies patterns, including unfair ones, from the data it learned from.

🔬 A real example: faces

In 2018, researchers Joy Buolamwini and Timnit Gebru tested commercial face-analysis systems. The systems were far less accurate for darker-skinned women than for lighter-skinned men. The training data had mostly lighter-skinned faces, so the systems learned them best.

📄 Another example: hiring

A large company built an AI to sort job applications, trained on CVs from people it had hired in the past, who were mostly men. The tool learned to rank women's CVs lower, and the company stopped using it. The AI faithfully copied an unfair history.

🗣️ Bias close to home

Voice assistants may understand some accents better than others. A credit-scoring app could learn to give fewer loans to people from certain areas if its data reflects past unfairness. In Kenya, where people speak many languages and live very different lives, data from one group may not represent everyone.

🛠️ Making AI fairer

Fairer AI needs varied data that represents everyone it will affect, testing accuracy separately for different groups, people reviewing important decisions, and openness about how a system decides. Asking "who could this hurt, and who is missing from the data?" is a skill everyone can practise.

Key points

• AI bias is when a system treats some groups unfairly.
• AI copies patterns, including unfair ones, from its training data.
• Face systems were once far less accurate for darker-skinned women.
• A hiring AI learned to prefer men because its past data did.
• Bias can affect accents, loans and much more.
• Fairer AI needs varied data, testing for every group, and human review.
• Always ask: who could this hurt, and who is missing from the data?

Try it for real: Bias detective

• Think of an AI that could be used in Kenya, such as a loan app, an exam marker or a voice assistant. Who might be missing from its data, and what could go wrong for them?
• Write one rule you would set to make it fairer.