Is It Legal to Reject a Resume on an AI-Detector Result?

What law actually applies when a detector flag decides a rejection?

No federal statute names AI detectors, and none has to. The moment a detector result decides who advances, it is a selection procedure under the Uniform Guidelines on Employee Selection Procedures, defined there as "any measure, combination of measures, or procedure used as a basis for any employment decision" ¹. From that point the ordinary rules apply.

Section 1607.3(A) is blunt: a selection procedure that has an adverse impact on the opportunities of any race, sex or ethnic group is discriminatory unless it has been validated ¹. The EEOC's guidance on tests and selection procedures states the standard the employer has to meet, job-related and consistent with business necessity, which it explains as "necessary to the safe and efficient performance of the job" ². It also names the second hurdle: even a job-related procedure loses if a challenger shows a less discriminatory alternative was available ².

That last clause is the one to sit with. If a cheaper procedure answers the same question and flags fewer people wrongly, the detector is the harder thing to defend, not the safer one.

New York City reaches the practice by name. Local Law 144 covers a computer-based tool that uses machine learning, statistical modeling, data analytics or artificial intelligence, helps an employer make an employment decision, and substantially assists or replaces discretionary decision-making ⁵. A classifier that sorts a resume into one bucket or another is doing exactly that, and DCWP is explicit that an employment decision includes screening rather than only the final hire ⁵. Where the law applies, a bias audit has to be completed before use, its summary posted publicly, and NYC candidates given notice 10 business days ahead; enforcement began 5 July 2023 ⁵.

New York City is not the only place writing rules for automated screening, so check the ones your postings reach. And read what a rejection reason has to survive before a candidate asks you for one.

Why a detector score is not a job-related criterion

Because it describes the document, not the work. Validation under the Guidelines connects a procedure to job performance, and nothing connects the statistical texture of a paragraph to how a compliance analyst reconciles a control or how a marketer briefs an agency. A vendor's published accuracy figure is not validity evidence, and the Guidelines say so in terms.

Section 1607.9 rules out the substitutes directly. The general reputation of a selection procedure, its author or its publisher, casual reports of its validity, all forms of promotional literature, and the testimonials and credentials of sellers are specifically excluded as evidence of validity ¹. A vendor's accuracy page is promotional literature. It may also be true, and it is still not the thing an enforcement agency asks to see.

The gap between accuracy and validity is where this whole question lives. Accuracy is the rate at which a tool agrees with a label on a benchmark corpus someone else assembled. Validity is documented evidence that what the procedure measures relates to performance on this job, for this applicant pool. No detector claims the second, because there is nothing to relate: the output is a property of the prose.

So before a detector goes near a requisition, ask the vendor for the second thing. The validation study, the job families it covers, the false-positive rate broken out by demographic group. The questions worth putting to a screening vendor are the ones your own counsel will put to you later.

How do detector false positives land, and on whom?

Not randomly. Seven widely used detectors were run over 91 human-written TOEFL essays and misclassified them as AI-generated at an average false-positive rate of 61.22 percent, with 97.80 percent flagged by at least one detector, while essays by US eighth-graders were classified with near-perfect accuracy ³. National origin is a protected class, so that error pattern is the adverse-impact case writing itself.

The follow-up experiment is the tell. When the same TOEFL essays were rewritten to enrich the word choices, the average false-positive rate fell from 61.22 percent to 11.77 percent ³. Simplifying the word choices in native-speaker essays pushed misclassification the other way ³. The tool is reading register, and register tracks first language, education and profession.

Then do the arithmetic on volume. Vanderbilt did it before disabling Turnitin's AI detector: against a claimed 1 percent false-positive rate and 75,000 submissions in a year, roughly 750 pieces of student work would have been wrongly flagged, and the university concluded the software was not an effective tool to use ⁶. One percent sounds like rounding until it meets a funnel. Whether detectors work at all is a separate question from whether their errors are survivable.

Title VII disparate impact requires no intent. A procedure that screens out non-native English writers at a materially higher rate is exposure whether or not anyone meant it, and the four-fifths rule is where the enforcement agencies start: a selection rate for any race, sex or ethnic group below 80 percent of the highest group's rate is generally regarded as evidence of adverse impact, and smaller gaps can still count where they are significant in statistical and practical terms ¹.

Does one detector threshold work across job families?

No. Baseline AI use at work differs sharply by occupation: highest in computer and mathematical, management, and business and finance jobs at 43, 42 and 41 percent, with every occupation group somewhere between 15 and 50 percent ⁴. Writing was the work task people most often used it for, at 39.5 percent ⁴. One threshold therefore carries a different false-positive burden into every family you apply it to.

Adverse impact is assessed per job, against the total selection process, and the agencies keep the discretion to reach an individual component of it where appropriate ¹. Running one detector across ten requisitions is not one exposure, then. It is ten, each with its own applicant pool and its own baseline rate of assisted drafting.

Register moves the same way. A compliance analyst writing to a controls template and a marketing coordinator writing to a brand voice do not produce prose with the same statistical texture before either of them opens an assistant. The same score means a different thing in each case, and a single cut line quietly sets a different standard for every job family it touches.

If a detector stays in the process, the minimum is per-family impact data. Records that disclose the impact a selection procedure has on employment opportunities by race, sex and ethnic group are already what the Guidelines ask users to maintain, with sampling permitted where applicant volume is large ¹. The adverse-impact arithmetic on a screening step is not hard to run. Not having run it is the finding.

What to run instead of the detector

Screen the claim, not the authorship. Pick two or three job-related criteria, write them down before you open the pile, and judge each submission against them. Where the real question is whether the person can do the thing, put a short piece of the work in front of them with AI allowed and the working visible. Authorship stops mattering once the work is watched.

• Write the criteria first. A rejection reason that existed before you saw the candidate is the one that survives being questioned.
• Ask about a decision in the work, not about the prose. Which figure did they check, what did they discard, what did the source turn out not to support. A model produces the paragraph; it does not produce the reason the paragraph says what it says.
• Move the weight to a work sample. Skipping the resume screen in favour of a short work sample converts an authorship question into an observation.
• If the detector stays, demote it. A flag routes a file to a human read and never rejects on its own. Record what the reviewer read and the job-related reason they gave, because "substantially assists" is the phrase that decides whether Local Law 144 applies ⁵, and a rubber stamp assists substantially.
• Tell candidates what you use. Notice under Local Law 144 runs 10 business days before use, not after a rejection ⁵.

None of this makes authorship answerable, and it is worth being straight that nothing does. Whether an AI-written resume should be disqualifying at all is a policy call to make explicitly. The alternative is making it implicitly, at a threshold a vendor chose.

See a sample report >