Fairness in AI — Equitable Outcomes and Non-Discrimination Requirements

Definition

Fairness in AI refers to the requirement that AI systems treat individuals and groups equitably, without producing discriminatory outcomes based on protected characteristics such as race, gender, age, disability, national origin, or religion. It is both a technical challenge — measuring and achieving equitable outcomes across subgroups in complex statistical systems — and a legal obligation under anti-discrimination law, GDPR, and the EU AI Act.

Fairness in AI is technically complex because no single, universally applicable definition of fairness exists. Multiple competing mathematical criteria exist: statistical parity (equal rates of positive outcomes across groups), equal opportunity (equal true positive rates across groups), predictive parity (equal positive predictive values across groups), and individual fairness (similar individuals should receive similar outcomes). These criteria are frequently mathematically incompatible — satisfying one often precludes satisfying another. Organizations must make explicit, documented choices about which fairness criteria apply to their specific use case and justify those choices.

ISO TR 24027 provides technical guidance on bias in AI and AI-aided decision making, including fairness evaluation methodologies. The EU AI Act requires that high-risk AI systems not produce discriminatory outputs and must be tested for non-discrimination before deployment.

Why it matters operationally

Fairness matters because unfair AI systems cause real harm at scale. A credit scoring model that systematically underestimates creditworthiness for specific demographic groups denies financial access to thousands of individuals through a single automated process. A hiring model that filters out qualified candidates from certain groups amplifies historical discrimination at the speed and scale of automation. A recidivism prediction system that overestimates risk for certain groups influences judicial decisions with life-altering consequences.

For organizations, unfair AI creates legal exposure under anti-discrimination law (independently of AI regulation), regulatory exposure under the EU AI Act, reputational damage when discriminatory outcomes become public, and increasingly, civil litigation from affected individuals. The scale of automated AI decisions means that a single fairness failure affects thousands of individuals before it is detected without adequate monitoring.

Regulatory framework

Framework	Fairness requirements
EU AI Act	High-risk systems must be evaluated for non-discrimination before deployment and monitored for fairness during operation. Data requirements (Article 10) include verifying that training data is free from discriminatory bias.
ISO TR 24027	Technical guidance on bias in AI: bias types, fairness evaluation methods, fairness criteria, and mitigation strategies.
GDPR — Art. 22	Automated decisions with significant effects on individuals cannot be based on sensitive characteristics without adequate safeguards.
Equal Treatment Directives (EU)	Algorithmic discrimination on protected characteristics violates equal treatment directives, regardless of whether it is produced by an AI system.

How Zertia evaluates it

Zertia evaluates AI fairness through two services. The AI Model Audit includes fairness testing as a core component: performance metric analysis across demographic subgroups, disparate impact assessment, and evaluation of whether fairness criteria are explicitly defined and appropriately chosen for the deployment context. The Ethical AI Mark evaluates conformity with ISO TR 24027 (Bias and Equity) as one of its four international ethics standards — providing independent certification that fairness controls meet international requirements.

[AI Model Audit] · Ethical AI Mark