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

Individual Fairness

Also known as: IF, individual-level fairness, similar treatment principle
Simply put

Individual fairness is an approach to fairness in machine learning built on the intuitive principle that similar individuals should be treated similarly by an automated decision system. It focuses on consistency at the level of the individual, rather than on outcomes compared across broader groups. As commonly framed, it is one of several competing conceptions of algorithmic fairness and is often contrasted with group fairness.

Formal definition

Individual fairness is a fairness criterion for ML-based decision processes requiring that individuals who are similar with respect to a given task be treated (or scored) similarly, thereby imposing a consistency constraint at the individual level. It is typically positioned as distinct from group fairness, which concerns comparisons of outcomes across defined groups; the two are conceptually separate and can be in tension rather than interchangeable. Note that individual fairness is a fairness concept, not the same as bias itself: it is one proposed way of correcting or constraining algorithmic bias, and its operationalization depends heavily on how 'similarity' between individuals is defined, which is contested and often application- or sector-specific (for example, healthcare). This entry does not resolve which fairness definition is appropriate for a given context, and it does not describe any binding legal requirement.

Why it matters

Individual fairness matters because much of the debate over algorithmic fairness has centered on group-level comparisons, yet decision subjects experience automated systems as individuals. The similar-treatment principle captures an intuition that many stakeholders find compelling: that two people who are alike with respect to the task at hand should receive comparable scores or outcomes, regardless of how their broader demographic group fares in aggregate. This gives compliance and model risk professionals a distinct lens for scrutinizing consistency, one that group-fairness metrics alone may not surface.

The concept is also important because it is frequently in tension with group fairness rather than reducible to it. As commonly framed in the literature, these are separate conceptions of fairness that can pull in different directions, so a system tuned to satisfy one may not satisfy the other. Teams that treat 'fairness' as a single, monolithic requirement risk optimizing for one criterion while assuming they have addressed all fairness concerns. Understanding individual fairness as one competing conception among several helps organizations make explicit, defensible choices about which conception applies in a given context.

Because individual fairness depends entirely on how 'similarity' between individuals is defined, its practical value is closely tied to a contested modeling decision. Defining a task-appropriate similarity measure is difficult and often sector-specific, and it can embed assumptions that themselves warrant scrutiny. This entry does not resolve which fairness definition is appropriate for any particular use, nor does it describe a binding legal requirement; it describes a conceptual criterion that professionals must operationalize carefully.

Who it's relevant to

Data scientists and ML engineers
For practitioners building and evaluating models, individual fairness introduces a criterion that hinges on defining a task-appropriate similarity measure between individuals. Because that definition is contested and often sector-specific, teams should document their similarity assumptions and recognize that satisfying individual fairness does not automatically satisfy group fairness, since the two conceptions can be in tension.
Model risk managers and validators
Individual fairness offers a distinct lens for probing consistency at the individual level, complementing group-based analyses. Validators should treat it as one competing fairness conception rather than a universal standard, and should scrutinize how 'similarity' is defined, since that choice materially affects conclusions about whether a system treats similar individuals similarly.
Compliance and policy specialists
Those advising on fairness obligations should note that individual fairness is a conceptual criterion, not a binding legal requirement described here. It is useful for framing internal fairness objectives, but professionals must decide which fairness conception applies in a given context, and this entry does not resolve that question.
Auditors and independent reviewers
Reviewers assessing fairness claims should confirm which conception of fairness a system targets, since individual and group fairness are conceptually separate and can conflict. A defensible audit examines both the similarity measure underlying any individual-fairness claim and whether that measure is appropriate to the domain and task.

Inside IF

Similarity-based principle
Individual fairness is commonly framed around the principle that similar individuals should receive similar model outcomes. It focuses on consistency of treatment at the level of the individual rather than parity across groups.
Similarity metric
Operationalizing individual fairness typically requires a task-specific distance or similarity measure that defines which individuals count as 'similar' for the decision at hand. Defining this metric is often the central and most contested design choice.
Lipschitz-style consistency condition
In many formal treatments, the constraint is expressed as a requirement that the difference in outcomes between two individuals be bounded relative to their measured distance, so small differences in inputs do not produce large differences in outcomes.
Relationship to group fairness
Individual fairness is a distinct fairness criterion from group-level notions (such as those defined over protected classes). The two can conflict, and satisfying one does not guarantee the other.
Distinction from bias measurement
Individual fairness concerns the fairness of treatment across comparable individuals, which is a normative criterion, and should not be conflated with bias as a statistical property of data or model outputs.

Common questions

Answers to the questions practitioners most commonly ask about IF.

Is individual fairness the same as group fairness?
No. Individual fairness, as commonly defined, focuses on the principle that similar individuals should receive similar treatment or outcomes from a model, whereas group fairness concerns statistical parity or comparable error rates across defined groups (for example, groups defined by protected attributes). The two are distinct and can be in tension: satisfying group-level parity does not guarantee that any two similar individuals are treated alike, and vice versa. Practitioners frequently err by treating a demonstrated group-fairness metric as evidence of individual fairness. They address different questions and generally require different measurement approaches.
Does achieving individual fairness mean the model is free of bias?
No. Individual fairness is one lens on fairness, not a guarantee that a model is unbiased. Bias and fairness are distinct concepts, and individual fairness typically depends on a chosen similarity measure between individuals; if that similarity measure itself embeds biased assumptions or omits relevant information, the model can satisfy an individual-fairness criterion while still producing outcomes many would regard as unfair. Individual fairness is a measure that can help reduce or surface certain fairness concerns, not one that eliminates bias.
How is similarity between individuals typically defined when applying individual fairness?
In many formulations, individual fairness relies on a task-specific similarity metric that quantifies how alike two individuals are with respect to the decision at hand. Defining this metric is one of the most contested and consequential steps, because the choice of features, distance function, and weighting directly determines which individuals are treated as 'similar.' There is no single authoritative similarity measure that applies across contexts; the appropriate definition is often domain-dependent and may itself require justification, documentation, and review.
How can individual fairness be tested or monitored in practice?
Approaches commonly involve examining whether individuals judged similar under the chosen similarity metric receive comparably similar model outputs, which may include constructing pairs or neighborhoods of similar cases and comparing their predictions or decisions. Because results depend heavily on the similarity metric and the data used, testing typically needs to document those assumptions explicitly. Monitoring over time may also be relevant, since changes in data or model behavior can affect whether the similarity-to-outcome relationship still holds. Specific methods vary by context and are still evolving.
Who within an organization is typically responsible for evaluating individual fairness?
Responsibilities often span multiple lines of defense: model developers or data science teams (commonly the first line) may design and implement the similarity metric and initial testing, while an independent review or validation function (often described as a second line) may assess whether the fairness approach is appropriate and adequately documented. Audit or a third line may provide additional independent assurance. The precise allocation depends on an organization's governance structure and, in regulated sectors, on applicable expectations. This reflects general practice rather than a universally mandated arrangement.
How does individual fairness relate to documentation and governance requirements?
Because individual fairness depends on subjective and consequential choices, particularly the similarity metric, governance practices commonly emphasize documenting the rationale for those choices, the fairness criterion selected, its limitations, and the testing performed. Such documentation can support review, challenge, and oversight. This is a matter of good governance practice in many frameworks rather than a single prescribed legal requirement; specific obligations depend on jurisdiction, sector, and the applicable regulatory or standards regime.

Common misconceptions

Achieving individual fairness automatically satisfies group fairness (or vice versa).
Individual fairness and group fairness are distinct criteria that can conflict in practice. A model may treat similar individuals consistently while still producing disparate outcomes across groups, and vice versa. Which criterion is appropriate depends on the context and cannot be assumed interchangeable.
Individual fairness is an objective, self-defining property of a model.
Individual fairness depends heavily on the chosen similarity metric, which encodes normative and task-specific judgments about which individuals should be treated alike. Different metrics can yield different conclusions, so the criterion is only as well-defined as the similarity measure behind it.
Enforcing individual fairness eliminates unfairness or bias from a system.
Individual fairness is a measure that can reduce or constrain certain inconsistent treatment; it does not remove all forms of unfairness or bias. Limitations in the similarity metric, the data, and the surrounding decision process can leave residual risks.

Best practices

Document the similarity metric explicitly, including its rationale, the features it relies on, and its known limitations, so reviewers can assess the normative choices embedded in it.
Evaluate individual fairness alongside relevant group fairness criteria rather than in isolation, and record where the two conflict so decision-makers can make trade-offs transparently.
Involve domain experts and, where appropriate, affected stakeholders when defining what 'similar' means for a given decision, since the metric encodes value judgments that are not purely technical.
Treat individual fairness as one control among several that reduces, rather than eliminates, unfairness risk, and describe its residual limitations in model documentation.
Distinguish individual fairness analysis from statistical bias testing in your validation records, keeping the normative criterion and the data property clearly separated.
Monitor individual fairness over time, since changes in data distributions or the deployed population can alter which individuals are effectively treated as similar.