Bias and fairness
Whether an AI system treats groups of people unfairly.
Bias can come from unrepresentative training data, flawed labels, proxy variables or design choices. Fairness is assessed with defined metrics across groups, before deployment and during monitoring.
On the exam
Removing a sensitive attribute isn’t enough: other variables can act as proxies. Look for testing and monitoring of outcomes across groups.
Related terms
- Model monitoringWatching a live model’s inputs, outputs and performance.
- ExplainabilityBeing able to say why an AI system produced a given output.
Test yourself with today’s question of the day.