COMPEL Glossary / fairness
Fairness
Fairness in AI is the principle that AI systems should produce equitable outcomes across different demographic groups and not perpetuate or amplify existing societal biases.
What this means in practice
Fairness is more complex than it appears because multiple mathematical definitions exist -- demographic parity, equalized odds, predictive parity, calibration -- and these definitions can conflict with each other. A model satisfying one fairness criterion may violate another. This means fairness is not a single property but a set of context-dependent choices that organizations must make explicitly for each AI application. Fairness engineering requires both technical tools (bias detection in training data, fairness-aware model design, disparate impact analysis) and governance processes (defining what 'fair' means in each context, ongoing monitoring for emergent bias, and clear accountability). In the COMPEL framework, fairness is assessed in Domain 15 (AI Ethics and Responsible AI).
Why it matters
Fairness in AI is not a single property but a set of context-dependent choices that organizations must make explicitly for each application. Multiple mathematical definitions exist and can conflict with each other, meaning an AI system satisfying one fairness criterion may violate another. Organizations that do not make deliberate, documented fairness choices risk both legal liability and reputational damage from biased AI outputs.
How COMPEL uses it
Fairness is assessed in Domain 15 (AI Ethics and Responsible AI) during Calibrate. During Model, fairness criteria are selected for each AI system based on context, regulatory requirements, and organizational values. The Evaluate stage measures fairness outcomes against chosen metrics. Fairness engineering requires both technical tools and governance processes, connecting the Technology pillar's capabilities with the Governance pillar's oversight structures.
Related Terms
Other glossary terms mentioned in this entry's definition and context.