Ethical AI

machine-learning statistics

Definition

Ethical AI

Ethical AI is the study of systematic bias, transparency, and accountability in automated decision-making systems. The primary objective is to ensure that machine learning models do not perpetuate discrimination against individuals based on protected attributes (e.g., race, gender) through various formal fairness criteria.

Statistical Fairness Criteria

Group Fairness: Requires specific performance metrics to be equal across protected groups. Common measures include:

• Demographic Parity: $\mathbb{P}(\hat{Y}=1|A=1)=\mathbb{P}(\hat{Y}=1|A=0)$.
• Calibration: $\mathbb{P}(Y=1|A,S=s)=s$.
• Error Rate Balance: Equalising False Positive and False Negative rates across groups.

Individual Fairness: Posits that similar individuals should be treated similarly. Formally, for a distance metric $d$ and model $M$: $D(M(\mathbf{x}),M({\mathbf{x}}^{\prime }))\le L\cdot d(\mathbf{x},{\mathbf{x}}^{\prime })$.

Fair Regression

In numerical prediction tasks, fairness is often achieved by identifying and adjusting for the influence of protected features.

Mitigation via Substitution: A common strategy involves training a standard regression model $y={\mathbf{w}}^{\top }\mathbf{x}+\beta s+b$ (where $s$ is the protected attribute) and then replacing $s$ with a constant $c$ during inference. Setting $c$ to the average value $\bar{s}$ of the privileged group ensures that predictions are debiased relative to the protected characteristic while preserving the influence of non-protected features.

Theoretical Constraints

Impossibility Theorem: Established by Kleinberg et al. (2017), it is mathematically impossible to satisfy calibration, false positive parity, and false negative parity simultaneously, except in cases of perfect prediction or identical group base rates. This necessitates explicit policy decisions regarding the distribution of error types.

Transparency and Accountability

Model Cards: Technical documentation detailing model architecture, training data characteristics, and performance disparities to ensure transparency and identify potential for disparate impact.