Ethical Considerations in AI: Bias and Fairness

Learning Objectives Define and differentiate between various types of bias, including selection, measurement, and algorithmic bias. Analyze a machine learning system's lifecycle to identify potential sources of bias in data collection, model training, and deployment. Calculate and interpret key fairness metrics, such as Demographic Parity and Equal Opportunity, using confusion matrices. Evaluate the concept of the fairness-accuracy trade-off in the context of system design. Propose and critique technical and procedural strategies for mitigating bias in AI systems. Articulate the societal impact of biased AI in real-world applications like hiring, loan approval, and criminal justice. If an AI hiring tool consistently rejects qualified female candidates, is the algorithm...

TermDefinitionExample Algorithmic BiasSystematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others.A facial recognition system that has a significantly higher error rate for women of color than for white men because its training data predominantly featured white male faces. Training Data BiasA type of bias that originates from the data used to train a model. This can include underrepresentation of certain groups, historical prejudices reflected in the data, or measurement errors.Training a loan approval model on historical loan data from an era where a bank systematically denied loans to minority applicants. The model learns this historical bias and perpetuates it. Protected AttributesSensitive characterist...

Demographic Parity (Statistical Parity) P(Ŷ = 1 | A = 0) = P(Ŷ = 1 | A = 1) Use this statistical test to check if the model's positive predictions (Ŷ=1) are distributed equally across different demographic groups (A=0 and A=1). A significant difference indicates a potential disparate impact. This metric focuses only on outcomes, not on whether the predictions were correct. Equal Opportunity P(Ŷ = 1 | Y = 1, A = 0) = P(Ŷ = 1 | Y = 1, A = 1) This metric checks if the True Positive Rate (TPR) is equal across different groups. It ensures that among all individuals who genuinely qualify for a positive outcome (Y=1), the model is equally likely to identify them correctly, regardless of their group (A). It's a measure of fairness for qualified candidates. The Fallacy...