Understanding AI Bias and Fairness
Where bias creeps into AI systems and practical ways to reduce it.
Inspired by learning material from Google
AI learns from data — and data reflects the world, including its inequalities. Understanding bias is key to building fair systems.
Where bias comes from
- Data bias: training data under-represents some groups.
- Labeling bias: human labels carry assumptions.
- Feedback loops: biased predictions shape future data.
Real consequences
Biased models can unfairly affect hiring, lending, healthcare, and more — harming real people.
How to reduce bias
1. Audit datasets for representation gaps. 2. Test performance across subgroups, not just overall. 3. Use fairness metrics and constraints. 4. Keep humans in the loop for high-stakes decisions.
Fairness is contextual
There are multiple, sometimes competing, definitions of fairness. The right one depends on the situation and stakeholders.
Key takeaways
- Bias enters through data, labels, and feedback loops.
- Always evaluate across subgroups.
- Fairness requires ongoing auditing and human judgment.