Defining bias in AI-systems: Biased models are fair models

TL;DR

This paper tackles the ambiguity of 'bias' in AI fairness by distinguishing a technical architectural term from a sociological concept. It traces the historical emergence of the bias term in neural networks, notably the bias input in $z = \sum_i w_i x_i + b$, and argues that architectural bias is not the same as statistical bias or societal discrimination. The authors contend that unbiasedness does not guarantee fairness, illustrating how sampling and distributional biases can yield unequal outcomes, and they propose a nuanced distinction between discriminating between (task-relevant) and Discrimination (harmful). By introducing concepts like Capital D Discrimination and Equity, the work advocates for equity-aware, socio-technical bias mitigation rather than purely equal-treatment strategies, aiming to align AI fairness with real-world harms and opportunities.

Abstract

The debate around bias in AI systems is central to discussions on algorithmic fairness. However, the term bias often lacks a clear definition, despite frequently being contrasted with fairness, implying that an unbiased model is inherently fair. In this paper, we challenge this assumption and argue that a precise conceptualization of bias is necessary to effectively address fairness concerns. Rather than viewing bias as inherently negative or unfair, we highlight the importance of distinguishing between bias and discrimination. We further explore how this shift in focus can foster a more constructive discourse within academic debates on fairness in AI systems.

Figures (3)

• Figure 1: Comparison of discussed early neuron models.
• Figure 2: Linearly separated scatter plot showing differentiation between cucumbers and apples.
• Figure 3: Flowchart illustrating discussed types of biases.