Historical bias:

• occurs when AI models are trained on data that reflects past societal prejudices or inequalities
• This bias perpetuates unfair outcomes
• requires auditing historical data for systemic inequities and adjusting datasets

Representation bias:

• occurs when certain classes or demographics are underrepresented in the training data
• results in a skewed model performance
• requires augmenting datasets with diverse samples or reweighting to balance representation

Measurement bias:

• comes from using inaccurate or inappropriate proxies for real-world phenomena
• distorts model predictions.
• requires the careful selection of features
• and validation against ground-truth data to ensure relevance

Algorithmic bias:

• emerges during the training or optimization process when the model’s design or learning algorithm introduces unfairness. * happens if optimization prioritizes accuracy for the majority group, neglecting minorities, or if the model overfits to biased patterns
• Mitigation involves incorporating fairness constraints, like adversarial training, during model development.