Individual fairness is the principle that an algorithmic decision system should treat similar individuals similarly: two people who are alike in every respect relevant to the task should receive similar predictions or outcomes. In AI ethics, it stands as the main alternative to group fairness, which compares statistics across demographic groups rather than outcomes for individual people.
The idea was formalized in 2012 by Cynthia Dwork and colleagues in the paper “Fairness Through Awareness”, which expressed it as a Lipschitz condition: the distance between two people’s outcomes must not exceed the distance between the people themselves, measured by a task-specific similarity metric. Everything about the concept, including its strength and its main weakness, depends on that metric.
How the Lipschitz definition of individual fairness works
The formal definition treats a classifier as a mapping from individuals to distributions over outcomes. For any two individuals x and y, a fair mapping M satisfies D(M(x), M(y)) <= d(x, y), where d is the similarity metric on individuals and D is a distance between outcome distributions, such as total variation distance.
The condition is a smoothness constraint: small differences between people may produce only small differences in treatment. It does not require identical outcomes for everyone, and it does not compare group averages. A model can satisfy it while still producing different base rates across groups, provided those differences reflect genuine differences under the metric.
| Component | Role | Who defines it |
|---|---|---|
| Similarity metric d | Measures how alike two individuals are for the task | Domain experts, regulators or society |
| Outcome distance D | Measures how differently two people are treated | Model designer |
| Lipschitz constant | Caps how fast outcomes may change with similarity | Model designer, often set to 1 |
| Randomized classifier | Allows outcome probabilities rather than hard labels | Model designer |
Dwork and colleagues stressed that the metric should be public and open to debate, because it encodes a moral judgment about which differences between people may legitimately justify different treatment.
Individual fairness versus group fairness
Individual and group fairness answer different moral questions, and neither implies the other in general. Group criteria such as demographic parity or equalized odds can be satisfied by a model that treats two nearly identical applicants very differently, while an individually fair model may still show large gaps between groups.
| Criterion | Unit of comparison | Main requirement | Main weakness |
|---|---|---|---|
| Individual fairness | Pairs of individuals | Similar people get similar outcomes | Requires a trustworthy similarity metric |
| Group fairness | Demographic groups | Statistics match across groups | Allows arbitrary unfairness within groups |
| Counterfactual fairness | One person in a causal model | Outcome unchanged if protected attribute changed | Requires a credible causal graph |
| Fairness through unawareness | Model inputs | Protected attribute is excluded | Fails under proxy discrimination |
A frequent misunderstanding equates individual fairness with fairness through unawareness. Removing a protected attribute does not make similar people receive similar outcomes, because proxy variables such as postal code can carry the same information. Individual fairness instead constrains the model output directly with respect to the chosen metric.
Approaches to the similarity metric problem
The central practical obstacle is that no universally correct similarity metric exists. Defining one is as hard as defining fairness itself for the task, and a biased metric simply transfers bias into a formally fair model. Research has therefore developed several ways to obtain or approximate the metric.
| Approach | How it works | Main limitation |
|---|---|---|
| Expert-specified metric | Domain experts weight relevant features | Subjective and hard to validate |
| Learned from human judgments | People label pairs as similar or not | Inherits annotator bias |
| Sensitive subspace removal | Distance ignores directions linked to protected attributes | Assumes those directions are identifiable |
| Representation learning | Embeddings learned to be fair and informative | Metric becomes opaque |
Training methods such as SenSR (sensitive subspace robustness) treat individual fairness as robustness to perturbations along sensitive directions, similar to adversarial training. The model is penalized when moving an input along a direction associated with, for example, gender changes the prediction substantially.
Measuring and enforcing individual fairness in practice
Practitioners usually test individual fairness through consistency checks rather than proving the Lipschitz condition across all inputs. The most common metric compares each prediction with the predictions for its nearest neighbors, and a high average disagreement signals a violation.
| Technique | Stage | Purpose |
|---|---|---|
| k-nearest-neighbor consistency | Evaluation | Detects similar people with divergent outcomes |
| Paired testing | Audit | Submits near-identical profiles differing in one attribute |
| Lipschitz regularization | Training | Penalizes sharp output changes between similar inputs |
| Post-processing projection | Deployment | Adjusts scores to satisfy the constraint |
Paired testing, borrowed from audit studies of housing and hiring discrimination, is especially useful in an AI audit because it requires no access to model internals. It also connects naturally to contestability: a person can ask why a near-identical applicant received a better outcome.
Limitations and open debates
Individual fairness is attractive because it matches the intuition of treating like cases alike, yet it carries real limitations that experienced practitioners keep in view.
- Metric dependence: the guarantee is only as fair as the metric, and the metric is a value judgment.
- Historical bias in features: if credit histories already reflect discrimination, people who look similar on paper may not be similarly deserving.
- Computational cost: verifying the condition for all pairs is infeasible for large populations, so checks rely on sampling.
- Compatibility with group goals: satisfying individual fairness does not guarantee parity, so organizations often need to address both.
These limits explain why many teams treat individual fairness as a complementary check alongside group metrics, and why the fairness-accuracy trade-off must be evaluated for each criterion separately.
The short version
Individual fairness means that if two people are alike in the ways that matter, they should be treated alike by the system. If two loan applicants have the same income, debts and payment history, a fair system should not approve one and reject the other.
The hard part is deciding what “alike in the ways that matter” means. That choice is a judgment about values, not a purely technical setting, so the definition of similarity must be open to scrutiny.
Where it makes a difference
A regional insurer uses a model to set premium risk scores for small businesses. Group metrics look acceptable: average scores differ little between owner demographics. An auditor then runs paired testing, submitting 400 pairs of profiles that are identical except for the owner’s first name and the neighborhood of the registered address.
In about one pair in nine, the scores differ by more than the insurer’s tolerance band, and the gap traces to a feature derived from neighborhood data. The team retrains with a Lipschitz penalty on that sensitive direction and removes the feature.
After retraining, the paired differences fall to the level of random noise while overall predictive accuracy drops only slightly. The case shows why group statistics alone can hide unfair treatment of individuals.