AI Glossary / Causal inference and experimentation / Intermediate

Disparate treatment

Definition

In machine-learning fairness discussions, disparate treatment refers to treating people differently through a decision process that uses sensitive group characteristics. The term also has legal meanings, including intentional differential treatment because of protected traits in U.S. employment law, which require context-specific legal analysis.

What it means

An algorithm can apply different rules or thresholds based on a person's group membership. That is distinct from a single apparently neutral rule producing unequal outcomes across groups, commonly discussed as disparate impact.
The technical glossary usage and legal doctrine should not be treated as perfectly interchangeable. A legal assessment considers the applicable jurisdiction, activity, facts and statutory requirements. Conversely, a technical review should not assume a system is fair merely because it omits a field labeled with a protected attribute.
Group information may also be collected for auditing rather than used to make individual decisions. Understanding where and why the information enters the system is essential to a careful description.

How it works

Trace the full decision process: data collection, feature construction, model scoring, thresholds, business rules and human overrides. Identify whether sensitive attributes directly affect how a person is evaluated, and distinguish decision use from separate monitoring or research use.
Compare relevant rules and test appropriate cases while considering whether apparent proxies or downstream processes introduce additional issues. Examine actual implementation rather than relying only on a feature list or policy statement. Statistical tests can reveal patterns and support investigation, but a single test does not settle intent, causal attribution or legal liability. Keep technical findings precise enough that reviewers can assess them in the proper context.

Three practical examples

• An audit discovers that a decision rule uses different cutoffs for otherwise comparable inputs depending on a sensitive group label.
• A review distinguishes demographic data held for fairness measurement from data passed into the production scoring rule.
• A team examines whether a human override process applies different standards to people from different protected groups.

Think of it this way

Think of two labeled lanes that apply different entry rules to otherwise comparable cases. The key issue is whether the group label changes the treatment. The analogy does not determine whether a particular real-world distinction is lawful or justified, and it does not capture every indirect or proxy-based form of discrimination.

Worked example

Consider a synthetic decision rule that accepts a score of at least 70 for group A but requires at least 80 for group B. Two otherwise identical test records both have score 75. The rule accepts the A record and rejects the B record solely because of the group-dependent threshold.
This clearly demonstrates direct group-based differential treatment in the toy implementation. It does not, by itself, establish a legal conclusion about a real organization, since the meaning of the groups and the applicable context are unspecified.
Now suppose both groups use threshold 70, but their acceptance rates differ. That difference raises another question about outcomes and the data-generating process. Equal thresholds do not guarantee equal outcomes or fairness, just as unequal observed rates alone do not reveal the exact mechanism behind them.

Illustrative case study

Fictional case: Magnolia Services reviews a model-assisted application workflow. The statistical model uses one scoring formula, but a later rules layer changes cutoffs using a sensitive group flag. The initial documentation had described only the model and omitted the rules layer.
The review team records the differential rule, its stated purpose and how it affects otherwise comparable cases. It separates the technical finding from the legal and policy assessment, which must consider the actual decision context. It also checks whether demographic data used for auditing are appropriately separated from operational decisions. No real organization or legal outcome is asserted in this example.

Common misconceptions and mistakes

Removing explicit sensitive fields does not guarantee fair treatment or eliminate proxy effects. Conversely, collecting demographic information for an audit is not the same operation as using it to decide an individual's outcome.
Disparate treatment and disparate impact are related but distinct concepts. A group outcome gap alone does not identify which mechanism produced it. Do not assume that a model's internal features capture the entire decision process: thresholds, rules and overrides matter. Technical fairness vocabulary should not be presented as a universal legal definition across countries and sectors.

Business use

The concept helps organizations review how group characteristics influence decisions and whether the documented process matches the implemented one. It is especially relevant when models are combined with business rules and human judgment.
Maintain a clear data-flow record showing where sensitive information is collected, transformed, used and retained. Test the complete workflow with suitable synthetic cases and examine real outcomes using appropriate access controls. Report specific findings, such as a group-dependent threshold, rather than an unexplained fairness label. Obtain context-specific legal and policy review for consequential decisions, while continuing technical investigation of proxies, outcome disparities and implementation errors.

People also ask

Is disparate treatment identical to disparate impact?

No. One focuses on differential treatment; the other concerns differential effects of a practice.

Does removing a sensitive column settle fairness?

No. Other variables and process steps can still matter.

Does every use of demographic data have the same meaning?

No. Auditing, research and individual decision-making are distinct uses that need separate assessment.

Check your understanding

A model uses the same formula for everyone, but a later rules engine applies group-specific thresholds. Should the review stop at the model's feature list?

Answer: No. The complete decision process includes the rules engine. The differential thresholds are a relevant technical finding that must be assessed in the applicable context.

Related terms

• Disparate impact
• Protected attribute
• Discrimination
• proxy variable
• algorithmic fairness
• decision threshold
• fairness audit
• human override

Learn next

Learn the distinction between decision rules and outcome disparities next. Trace one complete model-assisted workflow and identify every point where group information could affect treatment. Keep technical observations separate from conclusions that require jurisdiction-specific legal analysis.

More in Causal inference and experimentation

• D-optimal design - D-optimal design selects an experimental allocation that maximizes the determinant of its information matrix.
• Davies-Bouldin Index - The Davies-Bouldin Index is an internal clustering measure that compares within-cluster spread with separation between cluster centers.
• Detection error tradeoff - Detection error tradeoff describes the relationship between false alarms and missed detections as a binary detector's decision threshold changes.
• Dueling bandit - A dueling bandit is a sequential learning problem in which the learner chooses pairs of options and receives relative feedback, such as which option was...
• E-optimal design - E-optimal design chooses an experimental allocation that maximizes the smallest eigenvalue of its information matrix.
• Error rate - Error rate is the proportion of evaluated units counted as incorrect under a defined rule.

Sources

Google Machine Learning Glossary, disparate treatment: https://developers.google.com/machine-learning/glossary#disparate-treatment U.S. EEOC, Section 15 Race and Color Discrimination: https://www.eeoc.gov/laws/guidance/section-15-race-and-color-discrimination Official glossary and agency excerpts consulted. Legal discussion is limited to contextual terminology; jurisdiction-specific and final editorial review remain pending.

Previous: Disparate impact Next: Disparity