AI Fairness

Term: AI Fairness

AI Fairness is the property of an AI system that produces consistent, equitable outputs across different demographic groups, datasets and operating conditions, without systematically disadvantaging protected characteristics such as race, gender, age or disability.

Why it matters

Under the Equality Act 2010, UK organisations are liable for discriminatory outcomes regardless of whether the discrimination originates from a human decision or an automated system. An AI system that approves loans, screens candidates or triages patients at different rates across protected groups creates direct legal exposure.

The EU AI Act extends this through Article 10, which requires high-risk AI systems to use representative training data and demonstrate that outputs minimise discriminatory effects. For UK organisations serving EU customers or operating cross-border, both regimes apply simultaneously.

How it works in practice

Fairness is assessed across three measurable layers: statistical parity (equivalent outcome rates across groups), individual fairness (similar inputs producing similar outputs) and counterfactual fairness (outputs that do not change if a protected characteristic were altered). Each layer surfaces different failure modes. A recruitment model can pass statistical parity while failing counterfactual tests if it relies on proxy variables correlated with gender or ethnicity.

Fairness sits as the second dimension of our CIA+EFT Framework and is evaluated during every AI Behaviour Verification review.