Category: Trustworthy AI Principles

Fairness

Also known as: Algorithmic fairness, Model fairness

Simply put

Fairness is the quality of treating people impartially and without favoritism toward one group over another. In the context of AI systems, it generally concerns whether a model's processes and outcomes affect different individuals or groups in an even-handed and reasonable way. The term is commonly used but has no single agreed definition, and its meaning varies by context.

Formal definition

Fairness, as commonly defined in general usage, refers to impartial and just treatment—an absence of favoritism toward one side or another, spanning both process impartiality and outcome impartiality (a distinction familiar from the concept of organizational justice). Applied to AI and model contexts, fairness typically denotes the goal that a system's treatment of individuals or groups be equitable and defensible, though the evidence available here supplies only general-language and workplace/DEI definitions rather than technical AI fairness criteria. Fairness should not be conflated with bias: bias typically refers to systematic error or skew (statistical or societal) that may be measured, whereas fairness is a normative judgment about whether resulting treatment or outcomes are acceptable. Precise operational definitions of AI fairness (for example, specific statistical parity or error-rate criteria) are context-dependent, can be mutually incompatible, and are not established by the evidence provided.

Why it matters

Fairness sits at the center of debates about the acceptability of AI systems, because decisions that affect people's access to credit, employment, housing, benefits, or other opportunities are increasingly informed by models. When a system's treatment of individuals or groups is perceived as arbitrary or as favoring one side over another, it raises legal exposure, reputational harm, and ethical concern—regardless of whether the model performs well on aggregate accuracy metrics. Fairness is therefore a normative question about whether outcomes and processes are acceptable, not merely a technical performance question.

A recurring pitfall is that professionals treat fairness and bias as interchangeable. Bias typically refers to systematic error or skew that can be measured; fairness is a judgment about whether the resulting treatment is acceptable. A model can exhibit measurable bias yet still prompt disagreement over whether it is 'unfair,' and conversely a model with low measured bias may still produce outcomes that stakeholders regard as unjust. Because the term has no single agreed definition, teams that assume a shared meaning often discover late in a project that stakeholders, regulators, and affected communities were applying different standards.

The practical stakes are heightened by the fact that different formal notions of fairness can be mutually incompatible: satisfying one criterion may make another impossible to meet simultaneously. This means fairness cannot be fully delegated to a metric. Organizations that document which conception of fairness they are pursuing, and why, are better positioned to defend their choices, while those that leave the concept undefined risk both internal confusion and external challenge. Governance controls and measurement can reduce and manage fairness-related risk but do not eliminate the underlying normative disagreement.

Who it's relevant to

Data scientists and model developers

Practitioners building or tuning models must decide how, and whether, to translate a normative fairness goal into measurable criteria. Because different formal notions of fairness can be mutually incompatible, developers should recognize that no single metric settles the question and should distinguish measured bias from the broader judgment of whether treatment is acceptable.

Responsible AI, ethics, and policy specialists

These professionals define what fairness means for a given system and context, given that the term has no single agreed definition. They are typically responsible for reconciling process impartiality and outcome impartiality, and for documenting which conception of fairness the organization is pursuing and why.

Compliance officers and legal professionals

Fairness carries legal and reputational exposure where AI systems affect access to opportunities, resources, or support. Legal and compliance staff need to understand that fairness is a normative judgment distinct from measured bias, and that undocumented or undefined fairness standards are difficult to defend if challenged.

Auditors and model reviewers

Those examining a system should assess not only whether bias was measured but whether the organization defined its fairness objective, justified the criteria chosen, and acknowledged that alternative fairness notions may conflict. This entry does not supply specific technical criteria against which to audit; those must be established for the relevant context.

Inside Fairness

Fairness objective (normative goal)

Fairness refers to the normative goal that an AI or model system does not produce unjustified disparities in outcomes or treatment across individuals or groups. It is a value-laden objective rather than a single technical property, and its meaning depends on the social, legal, and application context.

Group fairness criteria

A family of statistical definitions that compare model outcomes across defined groups, such as demographic parity (similar selection rates across groups), equalized odds (similar error rates across groups), and predictive parity. These criteria are frequently mutually incompatible, so satisfying one can preclude satisfying another.

Individual fairness

The principle that similar individuals, as defined by a task-relevant similarity metric, should receive similar outcomes. This approach depends heavily on how similarity is defined and can be difficult to operationalize.

Relationship to bias

Bias, as commonly defined, refers to systematic error or skew in data, model behavior, or outcomes, whereas fairness is the broader normative judgment about whether resulting disparities are acceptable. Bias is a contributing factor that fairness assessments examine, but reducing measured bias does not by itself establish fairness.

Context and legal dimension

Fairness intersects with, but is not identical to, legal concepts such as non-discrimination and protected-attribute rules, which vary by jurisdiction and sector. What is legally permissible and what is considered fair are related but distinct questions.

Governance and risk management placement

Fairness typically appears both as an AI governance concern (policies, accountability, and oversight for acceptable outcomes) and as an input to model risk management (measuring and monitoring disparate performance as a source of risk). These roles overlap but should not be collapsed.

Common questions

Answers to the questions practitioners most commonly ask about Fairness.

Is fairness the same thing as the absence of bias in a model?

No. Bias and fairness are related but distinct concepts that experts are careful not to blur. Bias, as commonly defined, refers to systematic error or skew in data, model behavior, or outcomes, and is often a measurable, technical property. Fairness is a broader, normative concept concerned with whether the distribution of outcomes or treatment across individuals or groups is judged to be equitable or acceptable under a given set of values, policies, or legal standards. Reducing measurable bias may support fairness objectives, but a model can exhibit low measured bias on one metric and still be considered unfair under a different fairness definition, and vice versa.

Is there a single, authoritative definition of fairness that models must satisfy?

No. Fairness has contested and context-dependent definitions rather than one universally authoritative meaning. Different fairness criteria—such as those framed around equal treatment, equal outcomes, or calibration across groups—can be mathematically incompatible, meaning satisfying one may make it impossible to satisfy another simultaneously. What is regarded as fair also varies by jurisdiction, sector, application, and the values of affected stakeholders. As commonly understood, selecting a fairness definition is a normative and often policy or legal decision, not a purely technical one, and the appropriate choice should be documented and justified for the specific context.

How do teams typically choose which fairness definition to apply to a given model?

In many frameworks, the choice begins with the model's use case, the affected populations, the potential harms, and any applicable legal or policy constraints in the relevant jurisdiction. Because fairness criteria can be mutually incompatible, teams typically document the rationale for the metric or metrics selected, the trade-offs accepted, and the stakeholders consulted. Where legal standards govern a domain, those constraints often shape or narrow the acceptable definitions. This selection is generally treated as a decision requiring cross-functional input rather than one made solely by a technical team.

Where does responsibility for fairness sit across the lines of defense?

Fairness responsibilities are often distributed across governance structures rather than owned by a single function. In many organizations the first line—those who build and operate the model—implements fairness assessments and controls; a second line provides independent challenge, review, and oversight; and a third line, such as internal audit, evaluates whether the process is designed and operating as intended. The specific allocation depends on the organization's governance model. This distinction between AI governance roles (accountability and oversight structures) and the technical measurement of fairness should be maintained rather than collapsed.

How is fairness monitored after a model is deployed?

Fairness is typically not a one-time assessment. Because populations, data distributions, and usage patterns can change over time, many frameworks call for ongoing monitoring of the selected fairness metrics against defined thresholds or expectations. Monitoring outcomes can trigger review, escalation, or remediation. It is worth distinguishing this from monitoring for model performance degradation, which concerns predictive accuracy over time; a model may maintain performance while its fairness characteristics shift, or the reverse. Monitoring measures manage and surface fairness risk rather than eliminate it.

How should fairness assessments be documented for governance and review purposes?

Documentation commonly captures the fairness definition or metrics selected and the reasoning behind them, the groups and harms considered, the results of the assessment, the trade-offs accepted, any limitations of the analysis, and the residual concerns that remain after controls are applied. Recording what is out of scope—such as populations, attributes, or harms not evaluated—is also important, since it clarifies the boundaries of the assessment. Such documentation supports independent challenge, audit, and accountability, but it should describe fairness controls as measures that reduce or manage risk, not as steps that guarantee a fair outcome.

Common misconceptions

Fairness and bias are the same thing, so removing bias produces a fair model.

As commonly defined, bias is a systematic skew or error, while fairness is a normative judgment about whether outcomes are acceptable. Reducing measured bias can support fairness but does not by itself establish it, because fairness depends on context, chosen criteria, and applicable norms.

There is a single, universally correct fairness metric that a model can satisfy.

Multiple fairness definitions exist, and several group fairness criteria are mathematically incompatible in typical settings, meaning a model generally cannot satisfy all of them at once. Selecting a criterion is a context-dependent choice rather than a settled technical default.

A model that passes a fairness check is legally compliant and free of discrimination risk.

Fairness metrics and legal non-discrimination requirements are related but distinct, and legal obligations vary by jurisdiction and sector. A favorable metric result does not guarantee compliance, and controls reduce rather than eliminate the underlying risk.

Best practices

Define the fairness objective explicitly for the specific application before selecting metrics, documenting the affected groups, the relevant context, and the rationale for the chosen definition.

Select fairness criteria deliberately and acknowledge trade-offs, recognizing that group fairness definitions can be mutually incompatible and that the choice reflects a normative decision, not a purely technical one.

Keep fairness distinct from bias in analysis and documentation, examining bias as a contributing factor while separately assessing whether resulting disparities are justified.

Coordinate legal, compliance, and technical stakeholders so that fairness assessments account for jurisdiction- and sector-specific non-discrimination requirements without treating a metric result as a compliance conclusion.

Situate fairness in both governance and model risk management, using governance to set accountability and acceptable-outcome policies and model risk processes to measure and monitor disparate performance over time.

State the limitations of any fairness evaluation, including the metrics chosen, the groups and similarity assumptions used, and the residual risk that remains after controls are applied.