What Makes an AI Requirement in a Posting Defensible

Which law actually applies to a stated requirement?

The same law that applies to a test. Federal selection law is written around function, and the definition is wide: the Uniform Guidelines define a selection procedure as any measure, combination of measures, or procedure used as a basis for an employment decision, covering the full range of assessment techniques from paper-and-pencil tests and work experience requirements through informal or casual interviews and unscored application forms ¹. A posted qualification that removes applicants sits inside that definition by name.

The operative test sits in statute. Under 42 U.S.C. 2000e-2(k), once a complaining party shows a particular employment practice causes a disparate impact, the employer must demonstrate that the practice is job related for the position in question and consistent with business necessity, and the complaining party can also win by identifying a less discriminatory alternative the employer refused to adopt ². That is Title VII, so it reaches race, color, religion, sex and national origin. Age runs under the ADEA and disability under the ADA, with their own standards.

The principle underneath is older and blunter. Griggs v. Duke Power Co. held that Title VII proscribes not only overt discrimination but also practices that are fair in form but discriminatory in operation, that the touchstone is business necessity, and that a test must measure the person for the job and not the person in the abstract ³. Nothing there forbids requiring anything. It forbids giving a requirement controlling force without being able to show it is a reasonable measure of doing this job.

What a defensible AI requirement has behind it

Three artifacts, all of them written before the posting goes live. A task list naming the specific tasks in this role where an assistant is used. A written bar saying what clearing those tasks looks like in observable terms. And per-candidate evidence showing how each applicant was measured against that bar, produced the same way for everybody who reached that stage. Missing any one of them, the requirement is a preference with paperwork.

The first is the one teams skip, and it is the one that makes the other two mean anything. A bar copied from a competitor's posting was not derived from your work, so it cannot be defended by reference to your work. Sit with the person who does the job, list the tasks where an assistant genuinely gets used, and write the bar off that list. A guess about seniority will not survive the question of where the number came from.

The third is the one that decides a challenge. Evidence applied evenly is what a challenge actually turns on, and it has to exist per candidate. A policy saying everyone was measured the same way proves nothing about the person who was not hired. Two adjacent pieces cover the mechanics: writing the requirement so it is not legally vague, and setting a defensible bar for good enough at AI in a specific role.

One concrete illustration of what an unexamined cutoff can do sits in the EEOC's own files. Three tutoring companies paid $365,000 to settle claims that they programmed their application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older, turning away more than 200 qualified US applicants ⁴. That was a hard-coded rule, not a model, and a consent decree carries no admission of liability. The lesson transfers anyway: a criterion nobody examined, applied automatically, is the shape of the problem.

Where state law changes the answer

In several states, and not in the same direction. Illinois Public Act 103-0804 amended its Human Rights Act, effective January 1, 2026, making it a civil rights violation for an employer to use artificial intelligence that has the effect of discriminating on a protected basis in recruitment or hiring, to use zip codes as a proxy for a protected class, or to fail to give notice that AI is in use ⁵.

That is an effects standard. Texas wrote the mirror image: its Responsible Artificial Intelligence Governance Act, HB 149, also effective January 1, 2026, reaches only an AI system developed or deployed with the intent to unlawfully discriminate, and says in terms that a disparate impact alone does not establish that intent ⁷. Federal Title VII still applies in Texas on its own terms. Being compliant with AI hiring law is not a single coherent claim.

Both of those are aimed at the tools you screen with. It matters here because a stated requirement and an automated screen usually arrive together: the line goes into the posting, and the same line becomes a knockout question, a keyword filter or a ranking input inside the applicant tracking system. Once it does, both bodies of law are in play at once and the notice obligations attach to the tooling. In New York City that means a bias audit within the prior year, a posted summary of it, and notice to the candidate at least 10 business days before an automated employment decision tool is used ⁸.

Check the date on anything you read about this last year. The EEOC's two AI technical assistance documents, the ADA one from May 2022 and the Title VII adverse-impact one from May 2023, came off eeoc.gov in late January 2025 and are still gone; archived captures return normally while the live URLs return errors ⁶. Removing guidance is not repealing law. Title VII, the ADA, the ADEA and the Uniform Guidelines all still apply, but an article pointing at a live EEOC page for AI hiring guidance is pointing at nothing. State law turns over at the same speed: Colorado's 2024 AI act was repealed before it applied to anyone, and its replacement does not reach employers before January 2027 ⁹.

Put the AI in the duties when the paper trail is thin

If the three artifacts do not exist and nobody is going to write them this quarter, move the AI out of the qualifications block and into the description of the work. The responsibilities section carries no screening burden. It tells candidates exactly what the job involves, and it removes nobody from the pile, which means there is no showing to owe in the first place.

The duties column is not a dodge. Most teams are in exactly this position: they want candidates who work well with an assistant, they have not defined what that means, and they have not built anything that measures it. A duties line says the true thing. A requirement line says a thing the file cannot support.

The decision about which column to use is worked through in required or preferred, and the related question of what you may ask candidates once they are in the room is whether you can legally ask how they use AI. None of this is legal advice, and jurisdictions move fast enough that a posting reviewed a year ago is worth reading again. Have counsel look at the final wording and at whatever the applicant tracking system does with it.

Read the evidence >