The Less Discriminatory Alternative Is Usually a Stage, Not a Model

What counts as a less discriminatory alternative?

A different way of running the same selection that keeps the disparity lower and still serves the job-related purpose the procedure was there to serve. Both halves are load-bearing, and dropping the second one is the common mistake. Title VII has carried the alternative since the 1991 amendments: a plaintiff can establish disparate impact by identifying an alternative employment practice the employer refuses to adopt ¹.

The statute is where the second half comes from. Its disparate-impact provision says a practice is unlawful where a complaining party shows that a particular practice causes a disparate impact and the employer then fails to demonstrate that the challenged practice is job related for the position in question and consistent with business necessity, and separately where a complaining party names an alternative and the employer will not take it up ¹. The provision covers race, color, religion, sex and national origin; age and disability claims run under different statutes with different standards, which is one of several reasons this belongs in front of counsel before it goes in a policy.

The search for an alternative is supposed to start before any complaint, and that is the part the litigation-framed pages lose. Where two or more procedures serve the user's legitimate interest and are substantially equally valid for a given purpose, the Uniform Guidelines say the user should use the one demonstrated to have the lesser adverse impact, and they ask that a validity study include an investigation of suitable alternative selection procedures and suitable alternative methods of using the selection procedure ².

Read that last clause slowly, because it carries the whole practical argument: alternative methods of using the selection procedure. The search is not restricted to different products. It covers the cutoff, the knockout, the ranked list, the stage order and the weighting, all of which are ways of using what you already bought. Those Guidelines date from 1978 and are agency guidance rather than a statute, so treat the sentence as a design brief and check its current legal weight with counsel.

Where do you look first?

Split the impact ratio by stage before touching anything, because a single funnel-wide number cannot tell you which step produced it. Compute a selection rate for each stage: application to screen, screen to assessment, assessment to interview, interview to offer. One stage usually carries most of the gap, and it is often a stage nobody thinks of as a test.

Two things follow from the split. First, you learn whether the disparity came from a stage the vendor controls or one you configured, which is a different conversation with a different person. Second, you get a per-stage baseline, and without one you cannot tell later whether a change helped or simply moved the gap downstream.

Initial screening is the usual answer, and it is worth saying why. In the Harvard Business School and Accenture survey of 2,275 executives across the US, UK and Germany, more than 90% of employers who use a recruitment management system reported using it to filter or rank candidates at initial screening ³. That is employer self-report from early 2020, before generative AI reached hiring, and it describes employers rather than resumes. It still locates the stage worth opening first: the step doing the most cutting is the step most likely to be producing the ratio.

If the vendor holds the data you need, that is solvable, and the mechanics are in running an adverse impact audit when the vendor holds the data. If the worst stage turns out to be the resume screen, the prior question is whether that screen is throwing away people who would have done the job, because a stage that predicts nothing has no business necessity to defend in the first place.

Try the four cheap swaps before you change vendors

Four changes account for most of the disparity a hiring team can actually move, and none of them requires buying anything. Work them in this order inside the worst stage, changing one thing at a time so you can tell which change did what. Each is a way of using the same procedure differently, which is exactly what the Guidelines ask a user to investigate.

1. Soften or delete a knockout. A yes/no question that ends an application is the single most consequential setting in a funnel, because it removes people before any evidence is read. Ask what job requirement each one encodes. If nobody can write the requirement, the question goes. 2. Swap a ranked list for a pass bar. Ranking asks more of the evidence than screening does, and the Guidelines say so directly: evidence sufficient to support a procedure on a pass/fail basis may be insufficient to support the same procedure used on a ranking basis, and where ranking carries greater adverse impact than an appropriate pass/fail use, the user should have sufficient evidence of validity and utility to support ranking ⁴. 3. Move the cutoff to the proficiency the job needs. A shipped threshold is a product decision about pass rates. The standard in the Guidelines is that a cutoff score should normally be reasonable and consistent with normal expectations of acceptable proficiency within the work force ⁴, which is a claim about the job rather than about the tool. 4. Replace a proxy with the requirement behind it. School, postcode, employment gap and years since graduation are stand-ins for something. Write the underlying requirement, test for that, and the stand-in has nothing left to do.

The order matters more than it looks. Knockouts first, because they are binary and their effect is largest; the cutoff last, because moving it changes the volume every downstream stage sees, and a change made underneath a moving cutoff cannot be attributed to anything. Which of these is even available depends on what the tool exposes, and taking that inventory is its own exercise: most of the disparity comes from settings the employer chose.

Keep the record, and keep both numbers

Record two numbers for every change: the impact ratio and whatever predictive relationship you have. Lower the disparity by degrading the decision and you have a worse process wearing the alternative's clothes, so the case for keeping the change belongs to whoever proposed it. One shared document with a row per change is enough.

The trade-off is real and smaller than the older literature claimed. Modelling a six-method selection battery on the corrected 2022 estimates, Berry and colleagues found that zeroing the weight on cognitive ability tests moved the modelled adverse impact ratio from .42 to .67 while composite validity slipped only from .61 to .56, and that pushing the ratio past .80 cost more, down to .48. They say in the same paragraph that this does not fully resolve the trade-off, only that the reduction is less than previously thought ⁵. That is a modelling result under stated assumptions about the applicant pool and the selection ratio, not an outcome observed at any employer, and the six methods it covers do not include resume screening. Treat it as calibration.

The rejections are the valuable half of the record, because the question that comes back is why the obvious fix was rejected, and only the file answers that. California's fair-employment regulations, amended for automated-decision systems and effective October 1, 2025, make the point from the other direction: evidence, or the lack of evidence, of anti-bias testing is relevant to a discrimination claim and to any defence, including the quality, recency and scope of the effort, its results, and the response to the results ⁶⁷. That creates no duty to test and sets no standard for adequate testing, and it applies to California employers under that state's law. Testing and then doing nothing is worse than it looks. Confirm with counsel what that record should say for the jurisdictions you hire in.

On Monday, pick the single stage with the worst ratio, choose one threshold change inside it, and recompute both the ratio and the predictive relationship for that change alone before touching anything else. One change, two numbers, one line in the file.

Read the evidence >