Skip to main content
Bureau of AI
Category: Fairness & Bias

Nondiscrimination

Also known as: Non-discrimination
Simply put

Nondiscrimination is the principle of not treating people unfairly or excluding them because of characteristics such as race, color, national origin, or other protected traits. In practice, organizations adopt nondiscrimination policies committing to provide programs, services, and employment opportunities without such unfair treatment. As commonly used, it means the absence or avoidance of discrimination.

Formal definition

Nondiscrimination denotes the absence or avoidance of discrimination, typically operationalized through organizational policies and legal compliance obligations that prohibit adverse or exclusionary treatment of individuals on the basis of protected characteristics (for example, race, color, ethnic or national origin, and similar categories referenced in the source policies). In institutional contexts, it is frequently expressed as a commitment to comply with applicable state and federal civil rights laws and to ensure that no person is excluded from participation in, denied the benefits of, or subjected to discrimination under a program or service. The precise scope of protected categories, covered activities, and enforcement mechanisms is jurisdiction- and instrument-specific and is not defined uniformly across the sources provided; this entry does not address how the concept is applied to automated decision systems or AI-specific fairness testing, which are governed by separate and evolving standards not present in the evidence.

Why it matters

Nondiscrimination is a foundational civil rights principle that underpins how organizations design and deliver programs, services, and employment opportunities. The evidence shows public bodies and institutions—such as the Illinois EPA, the University of California, Merced, the Colorado Department of Health Care Policy and Financing, and NEMA—expressing formal commitments not to exclude, deny benefits to, or subject individuals to discrimination on the basis of protected characteristics like race, color, and ethnic or national origin. These commitments typically flow from obligations to comply with applicable state and federal civil rights laws, making nondiscrimination both a stated value and a compliance requirement.

For professionals in AI governance and model risk management, nondiscrimination matters because it is the legal and ethical backdrop against which fairness concerns in automated systems are often evaluated. However, the sources here address nondiscrimination as an organizational and legal principle, not as a technical fairness property of AI systems. It is important not to conflate a general nondiscrimination commitment with AI-specific fairness testing, bias measurement, or disparate-impact analysis of automated decisions—those are governed by separate and evolving standards not represented in this evidence. The precise scope of protected categories, covered activities, and enforcement mechanisms varies by jurisdiction and by the specific policy or law involved.

Because the definition and reach of nondiscrimination are instrument- and jurisdiction-specific, practitioners should treat any given organizational statement as a starting point rather than a universal standard. Reducing discrimination risk through such policies is a risk-management measure; it does not by itself guarantee that a program, service, or system is free from disparate treatment or impact in practice.

Who it's relevant to

Compliance officers
Nondiscrimination policies are typically expressed as commitments to comply with applicable state and federal civil rights laws. Compliance officers are responsible for ensuring that an organization's stated nondiscrimination commitments align with the specific legal obligations that apply to it, recognizing that scope and enforcement vary by jurisdiction and instrument.
Legal and policy specialists
The evidence shows nondiscrimination framed through formal policy statements and notices referencing protected characteristics such as race, color, and ethnic or national origin. Legal and policy professionals draft, interpret, and scope these commitments, and are best positioned to identify which protected categories, covered activities, and enforcement mechanisms apply in a given context.
AI governance professionals
Nondiscrimination serves as a legal and ethical backdrop for fairness considerations in AI oversight. Governance professionals should note that the sources here describe nondiscrimination as an organizational and legal principle, not as a technical fairness property of automated systems, and should avoid conflating a general nondiscrimination commitment with AI-specific fairness or bias testing, which are governed by separate and evolving standards.
Program and service administrators
Public bodies and institutions in the evidence commit to not excluding individuals from participation in, or denying them the benefits of, their programs and services. Administrators who deliver those programs are responsible for translating nondiscrimination commitments into day-to-day practice, while recognizing that a policy statement reduces but does not by itself eliminate the risk of discriminatory outcomes.

Inside Nondiscrimination

Protected characteristics
The attributes or group memberships that nondiscrimination principles are intended to safeguard, such as those defined by applicable anti-discrimination law in a given jurisdiction. The specific list of protected characteristics varies by legal regime and sector, so the applicable set should be determined from the governing law rather than assumed to be universal.
Disparate treatment vs. disparate impact
Two conceptually distinct forms of discrimination often referenced in nondiscrimination analysis. Disparate treatment typically refers to differing treatment on the basis of a protected characteristic, while disparate impact typically refers to facially neutral practices that produce differential outcomes across groups. Their precise legal meaning and burden of proof depend on jurisdiction and statute.
Bias and fairness as related but separate inputs
Bias, commonly understood as a systematic error or skew in data, model behavior, or outcomes, is distinct from fairness, which is a normative judgment about whether outcomes are acceptable. Nondiscrimination draws on assessments of both but is itself framed by legal and policy obligations rather than by any single technical fairness metric.
Outcome and process considerations
Nondiscrimination can be evaluated at the level of model outputs and decisions as well as at the level of the processes and data used to produce them. Both dimensions may be relevant depending on the framework or legal standard applied.
Documentation and evidence
Records that support the assessment of whether a system operates in a nondiscriminatory manner, which may include testing results, rationale for design choices, and monitoring outputs. The nature and sufficiency of required evidence is context-dependent.

Common questions

Answers to the questions practitioners most commonly ask about Nondiscrimination.

Does achieving fairness metrics mean a model is nondiscriminatory?
Not necessarily. Fairness metrics and nondiscrimination are related but distinct: satisfying a chosen statistical fairness measure does not, on its own, establish legal nondiscrimination, and different fairness metrics can conflict with one another. Nondiscrimination as a legal or regulatory concept typically depends on the applicable jurisdiction, the protected characteristics recognized there, and the legal theory involved (for example, disparate treatment versus disparate impact where such theories apply). A model can score well on a particular fairness metric while still producing outcomes that raise concerns under the relevant legal framework, and vice versa. Treat fairness metrics as one input to assessing nondiscrimination, not as a substitute for it.
Is removing protected attributes from the training data enough to prevent discrimination?
Commonly not. Excluding protected characteristics from inputs does not guarantee nondiscriminatory outcomes, because other features can act as proxies that correlate with protected attributes. This is often described as the problem of proxy discrimination or redundant encoding. In many frameworks, assessing nondiscrimination requires examining outcomes and potential proxy effects rather than relying solely on input exclusion. The appropriate approach depends on the applicable legal regime and sector; in some contexts retaining protected attributes for testing purposes may even be expected, while in others it is constrained. This entry does not resolve those jurisdiction-specific requirements.
How can teams test a model for potential discriminatory outcomes?
Testing approaches commonly include analyzing outcome disparities across relevant groups, examining features that may serve as proxies for protected characteristics, and applying one or more fairness or disparity measures suited to the use case. Because different measures can produce different conclusions, teams typically document which measures were selected and why. The specific metrics, thresholds, and protected characteristics that are relevant depend on the applicable legal framework, the deployment context, and organizational policy. This entry does not prescribe a universal testing methodology, as requirements vary by jurisdiction and sector.
Where does nondiscrimination testing fit within model risk management and AI governance?
Nondiscrimination considerations can appear in both AI governance and model risk management, though they play different roles. In AI governance, they typically inform policies, accountability structures, and oversight for how models are approved and monitored. In model risk management, discriminatory outcomes may be treated as a category of risk to be identified, measured, monitored, and controlled, often examined during validation and ongoing monitoring. The precise placement and required activities depend on the organization's framework and applicable guidance, and this entry does not assert a single mandatory arrangement.
Who is typically responsible for assessing nondiscrimination across the lines of defense?
Responsibilities are often distributed rather than held by a single function. In organizations using a lines-of-defense model, the first line (those developing or operating the model) may perform initial testing and controls, the second line (such as independent risk or compliance functions) may review and challenge those assessments, and the third line (internal audit) may provide independent assurance over the process. The exact allocation varies by organization, and legal counsel is frequently involved given the legal dimensions of discrimination. This entry does not specify a required structure, which depends on organizational design and applicable requirements.
How should nondiscrimination be monitored after a model is deployed?
Because outcomes and populations can change over time, nondiscrimination is commonly treated as an ongoing concern rather than a one-time check. Monitoring approaches may include periodic re-testing of outcome disparities, tracking changes in inputs or population characteristics, and revisiting proxy risks as data shifts. This is distinct from monitoring general model performance degradation, though the two can be examined together. The appropriate monitoring frequency, measures, and escalation triggers depend on the use case, applicable legal framework, and organizational policy, and are not fixed by this entry.

Common misconceptions

Removing protected characteristics from the input data guarantees nondiscrimination.
Excluding protected attributes does not by itself ensure nondiscriminatory outcomes, because other features can act as proxies and produce differential impacts. Nondiscrimination typically requires examining outcomes and impacts, not only the presence or absence of protected variables in inputs.
Satisfying a single technical fairness metric establishes nondiscrimination.
Fairness metrics reflect specific, sometimes mutually incompatible, definitions and do not automatically equate to legal nondiscrimination. Whether a system is nondiscriminatory depends on the applicable legal standard and context, which a single metric may not fully capture.
Nondiscrimination and bias mitigation are the same activity.
Bias mitigation is a technical effort to reduce systematic error or skew, while nondiscrimination is a broader obligation shaped by legal and normative standards. Reducing measured bias may support nondiscrimination but does not, on its own, demonstrate compliance with a nondiscrimination requirement.

Best practices

Identify the specific protected characteristics and legal standards that apply to your jurisdiction and sector before designing nondiscrimination assessments, rather than assuming a universal list or definition.
Assess both disparate treatment and disparate impact where relevant, and document which conception of discrimination each test is intended to address.
Test for proxy effects among non-protected features rather than relying on the removal of protected attributes to establish nondiscrimination.
Select fairness metrics deliberately, document why they were chosen, and acknowledge that they may be incompatible with one another and may not equate to legal nondiscrimination.
Evaluate nondiscrimination across both outcomes and the underlying data and processes, and monitor over time since impacts can shift as data and usage change.
Maintain documentation of design choices, testing, and monitoring sufficient to support the applicable legal or policy standard, treating these measures as risk-reducing rather than risk-eliminating.