AI Bias Explained: Where It Comes From and What to Do About It

AI systems can treat people unfairly even when nobody intends it. Understanding why helps you spot and reduce it.

• Training data: historical data reflects past inequalities and stereotypes.
• Missing data: some groups are under-represented, so the model performs worse for them.
• Labels: human annotators bring their own assumptions.
• Design choices: what the system is optimised for, and what it ignores.
• Proxies: neutral-looking features (like postcode) can stand in for sensitive traits.
• Feedback loops: biased outputs create data that reinforces the bias.
• Deployment: using a system in a context it wasn't built for.

• Recruitment tools that favour profiles resembling past hires.
• Face recognition with different error rates across groups.
• Image generators that default to stereotypes.
• Language models that associate certain jobs with certain genders.
• Credit or insurance models that disadvantage particular communities.

1. Define fairness for your use case. There's more than one measure.
2. Test across groups before and after launch.
3. Use representative data.
4. Keep humans in the loop for decisions that affect people.
5. Explain decisions and offer appeals.
6. Monitor over time. Bias can appear as data changes.
7. Include diverse voices in design and testing.
8. Document limitations and intended use.

• Question outputs that rely on stereotypes.
• Add specific, inclusive instructions to prompts, and review results.
• Report harmful outputs to the provider.
• Don't use AI as the sole judge of people.

Bias isn't a bug that can be fixed once. It needs ongoing testing, diverse input and human accountability.