Non-Discrimination

Non-discrimination in the context of AI regulation refers to the obligation that AI systems, their design, training data, deployment and outcomes must not produce unjustified differential treatment or disproportionate adverse impacts against individuals or groups on the basis of protected characteristics (e.g., race, gender, age, disability) or other relevant attributes. This concept is framed as both a substantive prohibition against discriminatory outcomes and a procedural duty to assess, mitigate and monitor risks of bias across the AI lifecycle. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))

Official / legal meaning: the European AI Act embeds non-discrimination among the fundamental-rights risks that AI regulation must address and repeatedly requires AI actors to avoid discriminatory outcomes when classifying, scoring, profiling or otherwise making decisions that affect persons’ rights and access to services. While the EU Act does not provide a single glossary entry titled “non-discrimination,” it expressly links non-discrimination to prohibitions on certain AI practices and to the classification and mandatory requirements for high-risk systems. In particular, recitals and operative provisions require AI systems used in contexts such as credit, employment, law enforcement, migration and social protection to be accurate, non-discriminatory and transparent so they do not reproduce or amplify historical patterns of disadvantage. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))

Jurisdictional variations:

• European Union: The EU AI Act treats non-discrimination as a fundamental-rights objective and operationalises it through risk classification (high-risk systems), mandatory requirements for high-risk AI (design, testing, documentation, post-market monitoring), and explicit recitals that tie discriminatory outcomes to breaches of Charter rights; enforcement is by market surveillance authorities and other competent bodies. The Act therefore emphasises outcome-focused assessment (disparate impact) as well as lifecycle controls. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))
• United States (federal): U.S. federal guidance (NIST AI RMF) treats fairness and “harmful bias” as a trustworthiness characteristic requiring measurement and management of bias. Executive Order 14110 directs federal agencies to guard against impermissible discrimination and coordinate enforcement of existing civil-rights laws; the FTC has signalled that discriminatory algorithmic outcomes can violate consumer-protection and sectoral statutes (e.g., ECOA) and urges pre-deployment testing and ongoing monitoring. The U.S. approach therefore blends voluntary technical risk-management (NIST) with enforcement under existing civil-rights and consumer-protection statutes. ([airc.nist.gov](https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/?utm_source=openai))
• International / standards bodies: OECD principles prioritize non-discrimination and fairness as part of "trustworthy AI" and call for safeguards and human-centred design. ISO/IEC terminology documents (e.g., 22989, related TRs) define related concepts (bias, fairness) and provide technical vocabulary for assessing systematic differences in treatment and ways bias can arise. UNESCO’s Recommendation frames non-discrimination as an ethical duty and links it to human-rights compliance and inclusive access. International instruments therefore provide normative principles and standardized terminology that regulators and businesses use to operationalise obligations. ([oecd.org](https://www.oecd.org/en/publications/oecd-employment-outlook-2023_08785bba-en/full-report/ensuring-trustworthy-artificial-intelligence-in-the-workplace-countries-policy-action_c01b9e49.html?utm_source=openai))

Context, scope and practical implications for businesses: Non-discrimination applies across the AI lifecycle — data collection and labelling, model design, validation, deployment, monitoring and redress. Practically, businesses operating across jurisdictions must: (i) identify contexts where decisions materially affect rights or opportunities (credit, hiring, housing, health, benefits, law enforcement), (ii) map applicable legal standards (EU fundamental-rights based rules; U.S. sectoral civil-rights and consumer-protection laws; sectoral/state rules such as California’s or Colorado’s AI-related laws), (iii) adopt technical and governance controls (impact assessments, representative datasets, fairness metrics, human oversight, documentation and logging), and (iv) implement monitoring, complaint-handling and corrective mechanisms. Failure to do so can trigger regulatory enforcement, civil liability, and reputational harm. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))

Key requirements / criteria (typical across frameworks):

• Contextual risk assessment: determine whether the AI use is high-risk or consequential and identify protected classes and vulnerable groups. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))
• Measurement of outcomes: use appropriate fairness metrics and disaggregated testing to detect disparate impact or disparate treatment. ([airc.nist.gov](https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/?utm_source=openai))
• Data and model governance: document data provenance, labelling protocols, known proxies for protected attributes, and model validation steps. ([khullani.github.io](https://khullani.github.io/AI-Governance-Handbook/?utm_source=openai))
• Mitigation and design choices: adopt proportional and technically feasible mitigations (reweighting, counterfactual testing, constraints, human review) and justify any differential treatment by objective, legitimate, and proportionate aims. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))
• Transparency and redress: provide meaningful disclosures, explainability to affected individuals, channels for contesting outcomes, and mechanisms for remediation. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))
• Ongoing monitoring: continuous post-deployment monitoring and documentation to capture emergent biases. ([airc.nist.gov](https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/?utm_source=openai))

Examples: An AI-driven credit scoring model that uses ZIP/postcode as an input may produce a disparate impact by reducing approvals for a racial minority; under EU rules that system may be high-risk and subject to mandatory design, testing and documentation; in the U.S. the same outcome could trigger enforcement under ECOA or FTC principles if it results in unlawful disparate impact. A hiring recommendation tool that systematically ranks candidates from certain age cohorts lower would require re-evaluation of training data, metrics and deployment constraints across jurisdictions. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))

Cross-references: See related concepts and obligations: bias, fairness, disparate impact, protected characteristics, algorithmic impact assessment, human oversight, and transparency / explainability. Authoritative sources for operational guidance include the EU AI Act (Regulation (EU) 2024/1689), the NIST AI RMF and its playbooks, OECD AI Principles, ISO/IEC terminology (22989) and UNESCO’s Recommendation on the Ethics of AI. ([eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng))