Algorithms are not neutral: Bias in collaborative filtering

Algorithms themselves introduce bias independent of data quality or developer intent. The paper challenges the common assumption that "clean data and good intentions" can eliminate algorithmic bias, demonstrating that the mathematical structure of algorithms like collaborative filtering inherently produces biased outcomes regardless of these factors.

Collaborative filtering systems exhibit "popularity bias" and "homogenizing bias" that marginalize already-vulnerable groups. These algorithms tend to recommend popular items and similar content, which systematically pushes recommendations away from niche items and diverse perspectives—effects that disproportionately harm people whose preferences fall outside mainstream distributions.

Statistical biases can translate into real discriminatory harm. The paper emphasizes that algorithmic biases are not merely mathematical quirks; they produce tangible discriminatory outcomes, particularly because "data points on the margins of distributions of human data tend to correspond to marginalized people."

Iterative learning creates a self-reinforcing bias cycle. As algorithms learn from user responses to their own recommendations, they create "selection bias in the course of learning," meaning the system's recommendations shape future data in ways that amplify initial biases over time.

The ubiquity of algorithmic decision-making makes this structural bias a serious societal concern. Given how widely these systems are deployed across platforms, the inherent biases built into algorithms warrant urgent attention and cannot be dismissed as edge cases.