Disparate impact vs disparate treatment in AI: definitions

Disparate treatment: definition and legal origin

Many federal antidiscrimination laws encompass two theories of liability. Disparate treatment is the first: intentional harm based on race, sex, disability or another proscribed motive. Regulators describe it the same way. New York's financial regulator, for example, defines it in lending as the intentional use of prohibited classifications, and notes that federal and New York law bar lenders from considering an applicant's sex and marital status in issuing credit.

Intent has legal weight. Constitutional discrimination claims under the Fifth and Fourteenth Amendments must assert intentional discrimination, and a disparate-impact showing alone is not enough. Under Title VII, a disparate-treatment claim allows damages in a way a disparate-impact finding does not.

Applied to AI, the definition points at what the system is built to consider. If a protected attribute is supplied to a model and used in producing a decision, the question is no longer about side effects. It is about the use of a prohibited classification. A related point from the NIST AI RMF Playbook is that some modeling algorithms or debiasing techniques that rely on demographic information could come into tension with legal prohibitions on disparate treatment. A well-intended fairness mitigation can therefore itself raise a disparate-treatment question.

Disparate impact: definition and legal origin

Disparate impact occurs when a seemingly neutral policy or action causes a disproportionate and unjustified negative effect on a group, regardless of intent. In a claim, a plaintiff challenges a facially neutral practice that has a disproportionately adverse effect along prohibited basis lines and is not justified by a legitimate rationale. Courts decide these claims using a burden-shifting effects test. The NIST AI RMF Playbook uses a similarly broad definition: a facially neutral policy or practice that disproportionately harms a group based on a protected trait.

Availability is statute-specific. The Supreme Court has held disparate-impact claims cognizable under the Fair Housing Act, but it has not held that they are available under every antidiscrimination statute. It reviews each provision, including whether the text contains effects-based language. NIST also observes that bias-measurement testing such as disparate impact analysis is not applied uniformly in the legal context.

As of October 2026, a Congressional Research Service product reports that federal disparate-impact enforcement is being deprioritized. That does not change the definition, and the theory continues to be described in legal sources under statutes such as the Fair Housing Act. Governance teams should track the position by statute and jurisdiction rather than assume the exposure has gone away.

Side-by-side: intent, mechanism and evidence

The two concepts differ along dimensions that determine how a policy, a control and an evidence package should be written. The table summarizes the differences relevant to AI systems.

• Intent: Disparate treatment concerns intentional use of a prohibited classification or motive. Disparate impact applies regardless of intent.
• Mechanism in AI: Treatment: a protected attribute, or demographic information, is used as an input or in mitigation. Impact: a facially neutral model or process produces disproportionate adverse outcomes for a protected group.
• Legal test: Treatment: whether a prohibited classification or motive was used. Impact: a burden-shifting effects test asking whether the practice has a disproportionate effect and is justified by a legitimate rationale.
• Unit of analysis: Treatment: an individual decision or interaction and its inputs. Impact: groups, compared across many outcomes.
• Evidence: Treatment: records of what the system received and used for individual decisions. Impact: statistical comparisons of outcomes across groups, such as outcomes for applicants with equivalent credit characteristics.
• Where it can be controlled: Treatment: at the input level, including in the execution path. Impact: through outcome-level testing and analysis outside the execution path.

How each surfaces in AI systems

Disparate treatment is a question about inputs and use. It can be examined for a single decision: what data reached the model, and whether a prohibited attribute was among it. An organization may reasonably choose to extend that input rule to attributes that directly stand in for a protected characteristic, but that is a policy decision to define explicitly, not a legal given.

Disparate impact cannot be seen in one decision. A single approval or denial reveals nothing about whether a neutral practice burdens a group disproportionately. That requires comparing outcomes across groups. NYDFS followed this approach in its 2021 Apple Card report, which analyzed disparate treatment and disparate impact separately and compared application outcomes for women and men with equivalent credit characteristics.

Measurement adds its own complexity. Group-fairness criteria can conflict. In the well-known COMPAS debate, one side pointed to unequal false-positive rates while the developers pointed to predictive parity. Later research found that not all fairness criteria can be satisfied when base rates differ across groups, and that disparate impact may result even when an instrument is fair on certain criteria. The metric and its justification are therefore a documented governance decision, not a default setting.

What each concept means for audit-ready evidence

Regulators examine the two theories separately, so evidence should be organized separately. For disparate treatment, an auditor needs to see what the system was permitted to receive, what it actually received in individual interactions, and what policy was applied. For disparate impact, an auditor needs the methodology: which populations were compared, which fairness metrics were selected and why, what thresholds applied, and what legal basis was used for the use case.

Two failure modes are common. The first is a single bias policy that uses both terms loosely, which leaves no owner and no testable requirement for either. The second is presenting evidence of one kind as an answer to the other: input controls do not establish the absence of disparate impact, and outcome statistics do not show that a protected attribute was never used.

Before deployment, mitigation techniques that use demographic information deserve legal review, because they sit at the point where the two theories meet. Policy should also name the AI use cases in scope and the statutes and regulators that apply to each, since the availability of disparate-impact claims is determined statute by statute.

Where runtime policy enforcement fits, and where it does not

Because disparate treatment turns on inputs in individual interactions, it is the concept that policy enforcement in the execution path can address. Trussed enforces policy inside the AI execution path across models, agents, tools and MCP servers, with actions to allow, redact, block or escalate for human approval. A rule can, for example, redact or block a protected attribute before it reaches a model, or route a case to a human reviewer. Enforcement at this point produces continuous audit-ready evidence of what policy was applied to each interaction, which supports a demonstration that protected attributes were not used as inputs.

The boundary matters. Runtime enforcement operates on individual interactions and tool calls. It does not measure whether outcomes differ across populations, and it does not replace the statistical analysis that disparate impact requires. That analysis is a separate discipline, run outside the execution path, with its own metrics, thresholds and owners. A regulated enterprise needs both: enforced input controls with evidence for treatment, and outcome measurement for impact.

Common questions

Short answers to questions teams raise when writing policy for these two concepts.

• {"question":"Can a model with no protected attributes as inputs still cause disparate impact?","answer":"Yes. Disparate impact is defined by the effect of a facially neutral practice on a group, regardless of intent. Excluding protected attributes addresses disparate treatment concerns about inputs, but outcomes must still be compared across groups to assess impact."}
• {"question":"Does using demographic data to reduce bias create legal risk?","answer":"It can. NIST notes that some modeling algorithms and debiasing techniques that rely on demographic information could come into tension with legal prohibitions on disparate treatment. Review such techniques with legal counsel before deployment."}
• {"question":"Are disparate-impact claims available under every antidiscrimination law?","answer":"No. They are cognizable under the Fair Housing Act, but the Supreme Court has not held that they are available under all antidiscrimination statutes. Each provision is reviewed, including for effects-based language."}