Chapter 1
Definition:
AI Bias is the systematic distortion in a model’s representation, measurement, or deployment, such that the outputs of an model that lacks accuracy, equity, or causes harm for certain groups.
Beyond “Is the model racist or sexist,” we move towards the question, “How may the design of this model cause disproportionate outcomes?”
Some examples of the bias we will explore includes:
Statistical Bias: systematic error in estimation or prediction. For AI models, this can result from sample selection bias, omitted data sets & variables, as well as aggregation bias (Simpson’s paradox).
Suppose you want to estimate the determinants of wages, but your wage data only includes people who are currently employed. You are then missing people who are out of the labor force/unemployed.
Social Bias
: Paterns in data or model behavior that reflect and reproduce unequal social hierarchies or stereotypes. This may include facial analysis and intersectional social bias, speech recognition, and ML discrimination.
Gender Shades: Commercial classification performed best on lighter-skinned men and worst on darker-skinned women, with substantially higher misclassification rates for the latter.
Normative bias
:The tendency for models to privilege or assumes a particular set of values or social ideals for the best outcome, justice, or decisions. This has been documented through Fairness and Abstraction in Sociotechnical Systems.
Ex) Hiring platform defines fairness as fundamentally a matter of prediction parity, rather than labor-market repair, accessibility, or anti-discrimination law.
Overview: Bias in AI Data Representation
-
Our downloads have everything you need to supplement this course.