What are the primary sources of bias in AI models as discussed in the chapter?

The primary sources of bias in AI models include historical data imbalances, flawed feature selection, and self-reinforcing feedback loops.

Bias in AI models primarily originates from three main sources: historical data imbalances, flawed feature selection, and feedback loops. Historical data often contains past discriminatory practices, such as redlining in mortgage lending, which AI models can replicate if trained on such data. Flawed feature selection can introduce bias when proxies for sensitive attributes, like ZIP codes correlating with race, are used. Feedback loops exacerbate bias by reinforcing existing disparities, such as when AI systems disproportionately flag transactions from certain demographics, leading to reduced access to financial services for those groups.