AI Glossary / Governance, ethics and safety / Intermediate
Individual fairness
Definition
Individual fairness is the principle that people who are similar in ways relevant to a particular task should receive similar treatment from a decision system. In machine learning, this is often expressed using a task-specific measure of similarity and a condition limiting how differently a model treats similar individuals. The similarity measure is central.
What it means
The focus is on comparisons between individuals, rather than only aggregate rates for groups. A model could select the same proportion of people from two groups while treating two similarly situated people very differently.
“Similar” does not mean identical in every recorded characteristic. It means similar in ways that can be justified for the decision being made. A convenient measure of distance in a dataset may include irrelevant differences or omit important circumstances.
This makes individual fairness both a mathematical and a contextual challenge. Once a similarity rule is chosen, a system can be checked against it. But satisfying that rule does not prove that the rule itself is appropriate, or that the wider process meets every fairness concern.
How it works
Define the task and what treatment means: a decision, score, ranking, or probability distribution over outcomes. Specify how similarity between individuals will be assessed for that task and justify the information used.
In the framework introduced by Dwork and colleagues, a classifier maps individuals to distributions over outcomes. A distance between those distributions is bounded by the task-specific distance between the individuals. Close individuals should therefore receive close outcome distributions.
In practical assessment, examine comparable cases, identify large unexplained differences, and investigate the data and decision logic involved. Review the similarity definition with relevant expertise and perspectives. Equal treatment under an unsuitable measure may simply formalize an inappropriate assumption. Group-level and individual-level analyses can provide complementary evidence rather than replacing one another automatically.
Three practical examples
1. A service checks whether two requests with the same relevant need and urgency receive comparable priority, despite differences in irrelevant account details.
2. A training-program selector examines whether people with comparable task-relevant preparation receive similar selection probabilities under its stated criteria.
3. A support-routing model compares otherwise equivalent cases submitted through different channels to investigate whether the channel produces unjustified differences in treatment.
Think of it this way
Think of using a measuring tape to check whether two pieces fit the same opening. The comparison is useful only if the measurement captures what fitting actually requires. Measuring color while ignoring width would miss the relevant similarity.
The analogy is limited because people cannot be reduced to interchangeable parts and the right criteria may be contested. It illustrates why selecting the measure is a substantive responsibility. A precise calculation cannot rescue a definition of similarity that fails to capture the task's legitimate considerations.
Worked example
Consider an invented binary allocation rule. The team defines a task-specific distance between two applicants as 0.05 on a scale from zero to one. It uses a condition requiring their allocation probabilities to differ by no more than that distance.
Applicant A receives probability 0.70 and applicant B receives 0.30. The absolute difference is 0.40, exceeding 0.05, so the pair fails the stated condition. A revised pair of probabilities, 0.52 and 0.48, differs by 0.04 and passes this particular check.
This does not prove the revised allocation is appropriate. The distance of 0.05 was an assumption in the example, and its justification matters. Other pairs may still fail, and the overall allocation may have other problems.
If the allocation is randomized, passing a probability-based condition also does not require identical realized outcomes on every run. Similar chances and identical individual outcomes are different requirements.
Illustrative case study
Fictional case: Beacon Courses reviews an AI system that recommends access to advanced workshops. Two participants with similar relevant preparation receive very different recommendations. The team discovers that an unrelated profile field strongly influences the score.
Reviewers examine whether the field has any justified connection to the workshop requirements and test comparable cases before revising the model. They also check whether the original preparation measure overlooks valid experience gained outside the organization's usual training route.
The case illustrates both sides of the problem: inconsistent treatment and an incomplete definition of similarity. It does not claim that removing one feature guarantees fairness or that a particular educational measure is appropriate in every setting.
Common misconceptions and mistakes
Individual fairness is not simply giving every person the same outcome. Relevant differences may justify different treatment, while irrelevant differences should not automatically do so.
It is also not guaranteed by equal group selection rates. Aggregate parity can conceal inconsistent treatment of particular pairs.
Another mistake is assuming that a distance generated by an embedding is automatically a valid fairness metric. Semantic or statistical proximity does not establish task-relevant similarity. Finally, a few favorable pair comparisons are not proof that the condition holds for all individuals in the intended population.
Business use
Businesses can use individual-fairness thinking to investigate complaints about inconsistent decisions and to review the criteria used in model design. Require a clear account of which differences matter for the task and why.
Comparable-case reviews can reveal issues missed by an aggregate dashboard, but they need reliable records and appropriate context. Document the similarity assumptions and their limitations, and allow them to be challenged when new evidence appears. The objective is a defensible pattern of treatment, rather than a mathematical label disconnected from how people experience the service.
People also ask
Is individual fairness the same as group fairness?
No. It focuses on comparable individuals; group fairness usually compares statistical properties across groups.
Who decides what makes people similar?
The definition requires task-specific justification and relevant expertise and perspectives. It should not be treated as an unquestionable default from the data.
Must similar people always get identical outcomes?
Not in every formulation. Some compare outcome probabilities or distributions, allowing similar chances rather than identical realized results.
Check your understanding
Two people have a defined task distance of 0.05, but their allocation probabilities differ by 0.40. Do they satisfy a rule that bounds the probability difference by that distance?
Answer: No. The difference exceeds the bound. The team should also review whether the distance itself is justified for the task.
Related terms
- Group fairness
- Demographic parity
- Counterfactual fairness
- Fairness constraint
- Individual Bias
- similarity metric
- decision consistency
- model evaluation
Learn next
Study similarity measures and group fairness next. Compare two hypothetical cases, explain which differences should matter, and test whether the proposed measure captures them without introducing irrelevant distinctions.
More in Governance, ethics and safety
- Incident response - Incident response is the coordinated work of investigating an incident, limiting its effects, addressing the problem, and restoring an acceptable operating...
- Incompatibility of fairness metrics - Incompatibility of fairness metrics means that some mathematical fairness conditions cannot all be satisfied simultaneously under particular data and...
- Independence - Independence, in probability, means that knowing one event or variable does not change the probability information about another.
- Inference optimization - Inference optimization improves how a trained machine-learning model runs when producing outputs.
- Information security - Information security is the protection of information and the systems that handle it against unauthorized access, disclosure, alteration, disruption, or...
- Inherited Bias - Inherited Bias arises when the output of one machine-learning system becomes input for another and carries an existing bias into the downstream learning...
Sources
Dwork and colleagues, Fairness Through Awareness: https://arxiv.org/abs/1104.3913
Google, Machine Learning Glossary, individual fairness: https://developers.google.com/machine-learning/glossary#individual-fairness
The original paper's framework and metric discussion, plus the glossary definition, were read. The probability calculation is an original simplified illustration of a specified condition, not a complete assessment method or claim of legal compliance.
Previous: Individual Bias Next: Individuals chart