Boards reviewing automated decisions often hear "adverse impact" and "disparate treatment" used as if they were interchangeable labels for unfairness. They are not. The distinction is older than machine learning, and it still governs how counsel frames risk, how auditors design tests, and how regulators characterize a failure. Getting the language wrong in a board pack does more than create awkward minutes. It can steer the organization toward the wrong evidence, the wrong remediation, and the wrong posture when an inquiry arrives.
This note is for directors and senior compliance counsel who need a working vocabulary they can defend. It is not legal advice, and it does not replace jurisdiction-specific counsel. It is a practitioner's account of how the two concepts show up in model review, and what a careful board should ask for when either term appears in a report.
Two different theories of harm
Disparate treatment, in the employment and civil rights tradition, concerns intentional differentiation on a protected basis. In plain terms, the decision maker treated people differently because of a protected characteristic, or applied a rule that was designed to produce that effect. Evidence often turns on policy text, training materials, override logs, communications, or design choices that encode an explicit preference. A model can be implicated when a feature, a label, or a hard-coded rule uses a protected attribute as an input, or when a proxy is used with the purpose of standing in for that attribute.
Adverse impact (often discussed alongside disparate impact) concerns facially neutral practices that produce unequal outcomes for protected groups, without requiring proof that the designer intended that result. The classic pattern is a selection rate, approval rate, or scoring threshold that systematically disadvantages one group relative to another, measured against a defined population and a defined decision. Statistical evidence is central. Intent may be irrelevant to whether the disparity exists, though intent and business justification still matter to how counsel evaluates legal exposure.
A board that collapses these ideas into a single "bias" finding loses the ability to ask the right follow-up. A finding of disparate treatment points toward design intent, documentation of prohibited inputs, and governance failures in how the system was built or instructed. A finding of adverse impact points toward population definition, metric choice, cutoff design, and whether a less discriminatory alternative was considered. Those are different workstreams, different owners, and different timelines.
How the distinction appears in model review
In an algorithmic review, disparate treatment questions are largely design and process questions. Did the system ingest protected attributes? If not, did it ingest near-proxies that the team knew or should have known would recreate those attributes? Were human overrides applied differently by group? Did prompt instructions, scoring rules, or vendor configuration language encode preferences that a regulator would read as intentional differentiation?
Adverse impact questions are measurement questions. Given the actual population subject to the decision, what are the selection or scoring rates by required category? What impact ratios follow from those rates? Do the disparities persist after reasonable controls that the business claims are job-related or credit-related? Does a material change to the model, the threshold, or the downstream rule change the ratios in a way that matters?
NYC Local Law 144, for example, is built around measured disparate impact style metrics for covered automated employment decision tools: selection rates and impact ratios by prescribed categories, produced through an independent bias audit, with publication and notice obligations attached. That regime is not a substitute for a disparate treatment analysis. An employer can publish clean impact ratios and still face a treatment theory if the tool or the surrounding process encodes intentional differentiation. Conversely, a team that proves it never used a protected attribute as an input has not thereby shown that adverse impact is absent.
Federal employment discrimination doctrine has long recognized both theories. Boards should not assume that "we removed race and sex from the feature set" closes the adverse impact conversation, or that a favorable impact ratio closes a treatment conversation. Both claims remain available in the broader legal environment surrounding hiring, credit, housing, and similar decisions, depending on the facts and the jurisdiction.
Language that helps the board stay precise
When a paper reaches the board, the useful discipline is to force every fairness claim into one of three buckets: treatment risk, impact risk, or residual uncertainty that has not yet been classified.
Treatment risk language should name the mechanism. "The resume screen includes a hard rule excluding candidates from a named set of postal codes known to correlate with a protected class" is a treatment-shaped statement. "The model may be biased" is not. Impact risk language should name the population, the decision, the metric, and the magnitude. "Among applicants in the audited window, the selection rate for group A was X and for group B was Y, producing an impact ratio of Z" is an impact-shaped statement. Vague severity adjectives without numbers are usually a signal that the measurement work is incomplete.
Boards should also resist the marketing habit of calling every fairness review a "bias audit." In regulated contexts, that phrase often implies a specific measurement exercise with independence requirements and prescribed outputs. A design review for prohibited inputs is valuable. It is not the same deliverable. Mixing the labels invites false comfort.
A short glossary for board materials helps. Use "disparate treatment risk" when discussing intentional or explicitly encoded differentiation. Use "adverse impact" or "disparate impact risk" when discussing measured outcome disparities under a facially neutral process. Use "model risk" for the broader governance topic that includes accuracy, stability, and operational failure, of which fairness is one component. Consistency across audit, legal, and product papers matters more than elegance.
What directors should ask next
If the paper alleges disparate treatment, the board's natural questions concern ownership and containment. Who approved the rule or feature? Has the prohibited mechanism been disabled in production? Is there a contemporaneous record of when it was introduced and when it was removed? Has counsel assessed disclosure and remediation obligations?
If the paper alleges adverse impact, the board's natural questions concern measurement integrity and alternatives. Which population and decision were measured? Were the categories complete enough for the claimed conclusion? Was the sample period representative of current use? What less discriminatory alternatives were tested, and who decided not to adopt them? Is the finding point-in-time, or is monitoring continuing on a known cadence?
If the paper is silent on which theory it means, the board should send it back. Ambiguity at this level is not a semantic inconvenience. It is a control failure in how risk is described to the people charged with overseeing it.
A practical posture for board materials
The organizations that handle this cleanly tend to do a few ordinary things well. Their model inventory tags high-stakes decisions and the jurisdictions that attach to them. Their audit and assessment workstreams separate design reviews (treatment-oriented) from outcome measurement (impact-oriented), even when the same engagement produces both. Their board summaries use the two terms with the same meanings counsel would use in a letter. Their remediation trackers distinguish fixes that remove an intentional mechanism from fixes that change a threshold, a feature set, or a process to reduce measured disparity.
None of that guarantees a favorable legal outcome. It does mean that when a director asks whether the company faces a treatment problem or an impact problem, the answer can be specific, evidenced, and consistent across the room. That is the standard the vocabulary is for. The models will keep changing. The distinction between treating people differently on purpose and producing unequal outcomes through a neutral-looking process will not.