AI biases as asymmetries: a review to guide practice

Area of Science:

• Artificial Intelligence
• Computer Science
• Ethics in AI

Background:

• Traditional views of bias in AI systems as mere errors or flaws are being challenged.
• Emerging perspectives recognize bias as an inherent and sometimes beneficial aspect of AI.

Purpose of the Study:

• To re-evaluate the understanding and measurement of bias in AI systems.
• To provide guidance on which AI biases to accept or amplify and which to minimize.
• To address the evolving landscape of AI bias and its implications.

Main Methods:

• Reviewing the paradigm shift in understanding AI bias.
• Defining bias as "violations of a symmetry standard" (following Kelly).
• Categorizing problematic biases into erroneous representation, unfair treatment, and violation of process ideals.

Main Results:

• Many AI biases, under the symmetry standard definition, are benign.
• Identified three key categories of undesirable AI bias: erroneous representation, unfair treatment, and violation of process ideals.
• Highlighted potential points of bias occurrence throughout the AI development and application pipeline.

Conclusions:

• A new framework for understanding and measuring AI bias is proposed, focusing on symmetry violations.
• Guidance is offered for differentiating between acceptable and undesirable biases in AI.
• The research provides a nuanced approach to managing bias in AI development and deployment.