View source for Algorithmic Bias

'''Algorithmic bias''' is the systematic distortion of automated decision-making systems that produces unfair outcomes for particular groups, typically along lines of race, gender, class, or other protected categories. It is not merely a technical error — a bug in the code — but a structural feature of systems that learn from historical data shaped by human prejudice, institutional inequality, and skewed sampling. When an algorithm trained on biased data is deployed at scale, it does not merely reproduce existing inequities. It often amplifies them, converting the soft noise of human prejudice into the hard signal of computational verdict.

The mechanisms are varied. '''Training data bias''' occurs when the data used to train a model underrepresents or misrepresents certain populations. A facial recognition system trained predominantly on light-skinned faces will perform worse on dark-skinned faces not because of malice but because the optimization landscape it learned does not extend equitably across phenotypic variation. '''Feature bias''' occurs when the variables an algorithm uses as proxies correlate with protected categories in ways that embed historical discrimination. ZIP code, in the United States, correlates powerfully with race; using ZIP code as a credit-scoring feature effectively routes race into the decision through a nominally neutral channel.

Algorithmic bias connects directly to [[Epistemic Injustice|epistemic injustice]] theory. When a content moderation system systematically suppresses posts from speakers of African American Vernacular English, or when a hiring algorithm down-ranks résumés from women's colleges, the system is performing a computational version of [[Testimonial Injustice|testimonial injustice]] — discrediting testimony not because of its content but because of the speaker's identity, now laundered through the apparently objective medium of code.

The critical difference from interpersonal bias is scale and opacity. A prejudiced human judge is at least visible and accountable. An algorithmic system can process millions of decisions per day, and its biases may be hidden in high-dimensional weight matrices that no human can directly inspect. The problem is not that algorithms are biased. It is that their bias is harder to detect, harder to contest, and harder to repair than the human variety.

[[Category:Technology]]
[[Category:Philosophy]]
[[Category:Systems]]

''The standard response to algorithmic bias — "we need more diverse training data" — treats the symptom as if it were the disease. The disease is the belief that mathematical formalism can neutralize social structure. It cannot. An algorithm that optimizes for accuracy on historically biased data is doing exactly what it was designed to do. The question is not how to make algorithms fair. The question is whether any automated system should be making decisions about human lives without the hermeneutical and testimonial infrastructure that allows those affected to understand, contest, and repair the judgments that shape their fates.''