Title VII disparate impact and AI hiring tools: what still applies

Title VII of the Civil Rights Act makes an employer liable when a neutral hiring practice, including an AI resume screener, assessment or interview score, excludes people of a particular race, color, religion, sex or national origin at a disproportionate rate and the employer cannot show the practice is job related and consistent with business necessity. That rule is written into the statute, and it applies to algorithms the same way it applies to paper tests. As of September 2026 the EEOC has said it will not pursue disparate impact cases itself, but private lawsuits and state laws still can.

This page covers the legal test, how to run the four-fifths calculation on your own tool's output, what the 2025 executive order and the EEOC's 2026 enforcement plan did and did not change, and a record you can keep for each tool. For the state and city AI statutes that sit on top of Title VII, see AI hiring laws by state; this page is only about the federal discrimination theory underneath them.

The legal test in three steps

The Supreme Court recognized disparate impact in Griggs v. Duke Power Co. (1971), where a diploma requirement and aptitude tests screened out Black applicants without being shown to predict job performance. Congress wrote the burden of proof into the statute in 1991, in 42 U.S.C. § 2000e-2(k). Commonly summarized, it runs in three steps:

1. The candidate shows a disparity. A particular employment practice causes a disparate impact on a protected group. The statute says each challenged practice must be identified, unless the parts of the decision process cannot be separated, in which case the whole process can be analyzed as one practice.
2. The employer shows job relatedness. The practice is "job related for the position in question and consistent with business necessity."
3. The candidate shows a less discriminatory alternative. Even if the practice is job related, the employer is liable if an alternative with less impact was available and the employer refused to adopt it.

Intent is not part of the test. An employer that never looked at demographic data, and whose vendor promised the model was fair, is in the same position as one that knew. That is why the theory matters so much for automated tools: the disparity usually comes from training data or proxies nobody chose on purpose.

One more line in the statute catches people out. Section 2000e-2(l) makes it unlawful to adjust scores, use different cutoff scores, or otherwise alter test results on the basis of race, color, religion, sex or national origin. So "fixing" a biased tool by re-weighting its output by group is not a lawful fix. The fix has to be in the procedure itself.

Where AI tools fit: what counts as a selection procedure

The Uniform Guidelines on Employee Selection Procedures (29 C.F.R. part 1607) define a selection procedure broadly: any measure, combination of measures or procedure used as a basis for an employment decision. They were written in 1978 for tests and interviews, but nothing in the definition depends on whether a human or software produces the result.

The borderline cases are where the risk hides. A sourcing tool that decides which profiles a recruiter ever sees is filtering. A transcript that is less accurate for some accents can produce thinner evidence and lower ratings for those candidates even though a person made the call. Test what reaches the decision, not only the part labeled "AI".

Running the four-fifths check on your own numbers

Under 29 C.F.R. § 1607.4(D), a selection rate for any race, sex or ethnic group that is less than four-fifths (80%) of the rate for the group with the highest rate will generally be regarded by the federal enforcement agencies as evidence of adverse impact. The same section says smaller differences can still be adverse impact where they are statistically and practically significant, and larger differences may not be where the numbers are small and not statistically significant.

An invented example: an AI screen reviews 400 applications for a customer support role.

Group B's ratio is under 80%, so under the Guidelines' rule of thumb this is evidence of adverse impact at the screening step, even though the overall hiring numbers at the end of the process might look balanced. Run the check at each step a tool controls, not only on final hires; the statute lets a challenger point at a particular practice.

Practical notes on doing this properly:

• Where the demographic data comes from. Most employers use voluntary self-identification collected separately from the application and kept away from the people making decisions. Do not ask interviewers to guess.
• Small numbers. With a few dozen applicants, one person changes the ratio by several points. Pool several months, or ask someone qualified to run a significance test, before drawing conclusions.
• How often. For employers subject to the full recordkeeping rules, § 1607.15 says adverse impact determinations should be made at least annually for each group that is at least 2% of the relevant labor force. Checking whenever the vendor ships a new model version is sensible practice on top of that.
• What to do with a failing ratio. Stop and look: is the step job related, is there validation evidence, and is there an alternative with less impact? § 1607.3(B) says that where two procedures are substantially equally valid, the one shown to have lesser adverse impact should be used.

What changed in 2025 and 2026, and what did not

The executive order

Executive Order 14281, signed April 23, 2025, states a policy "to eliminate the use of disparate-impact liability in all contexts to the maximum degree possible," directs agencies to deprioritize enforcement of statutes and regulations to the extent they include disparate impact liability, told the Attorney General and the EEOC Chair to review pending investigations and suits relying on the theory, and asked the Attorney General to consider whether federal law preempts state laws imposing disparate impact liability. We did not find a published preemption determination as of September 2026.

The EEOC's National Enforcement Plan

On June 4, 2026, the EEOC released a National Enforcement Plan replacing its previous strategic enforcement plan. The signed plan says the Commission will prioritize disparate treatment theories, will eliminate the use of disparate impact theories in investigations "to the maximum degree possible," and "will not commence, develop, or continue to pursue litigation advancing disparate impact claims." The same plan lists claims about the interpretation of disparate impact under the Civil Rights Act of 1991, the ADEA and the ADA among the legal questions it wants developed. Separately, the EEOC removed its 2022 and 2023 technical assistance documents on AI from its website in 2025.

What stayed the same

The practical reading: federal agency enforcement of disparate impact is unlikely in the near term, but the exposure has moved to private class and collective actions and to states. The best-known private case, Mobley v. Workday in the Northern District of California, has been proceeding on disparate impact theories against a screening software vendor, including an age discrimination collective the court allowed to go forward in 2025. It was still pending as of September 2026; read the current docket rather than a summary, including this one, before drawing conclusions from it.

Age and disability run on different rules

Title VII covers race, color, religion, sex and national origin. Two other federal statutes matter for AI screening and work differently:

• Age (ADEA). Disparate impact is available under the ADEA for people 40 and over, but the employer's defense is that the practice is based on a reasonable factor other than age (29 U.S.C. § 623(f)(1)), which is a different and generally easier standard than business necessity. Whether rejected applicants, as opposed to employees, can bring ADEA disparate impact claims has been contested in the courts. Tools that reward recent graduation dates, "digital native" language or years-of-experience caps are the usual sources of age disparity.
• Disability (ADA). The ADA prohibits qualification standards and tests that screen out people with disabilities unless job related and consistent with business necessity, and treats failing to make reasonable accommodation as discrimination (42 U.S.C. § 12112(b)(5) and (b)(6)). A timed game, a video analysis of eye contact or a speech-rate measure can screen out a qualified person who needed an accommodation they were never offered. See ADA accommodations in interviews.

Questions to put to every AI hiring vendor

The employer uses the procedure and makes the decision, so the employer carries the Title VII exposure, whatever the contract says about indemnities. Get these answers in writing before the tool touches a live requisition:

1. What exactly does the tool output: a pass or fail, a score, a ranking, a summary, or a transcript?
2. Which job-related criteria does it measure, and how were they tied to this role's actual duties?
3. What adverse impact testing has been done, on what population, by whom, and when? Can we see the selection rates by group, not only a statement that it "passed"?
4. Will you give us the data we need to run our own four-fifths check on our own applicants?
5. What validation evidence exists that the scores predict performance in jobs like ours?
6. Which inputs could act as proxies: zip code, graduation year, employment gaps, school names, voice or video features?
7. How do candidates request an accommodation or an alternative process, and does the tool flag when one was used?
8. What changes when you release a new model version, and will you tell us before it goes live on our account?
9. Can a recruiter see why a candidate received a given result, in terms a person can check?

A disparate impact record for each tool

Keep one of these per tool, per step it controls. It is also most of what the state notice and audit laws ask for, which saves writing it twice.

DISPARATE IMPACT RECORD — [tool] — [hiring step] — reviewed [date]

What the tool does at this step:
  Output:                 [pass/fail / score / rank / summary]
  Used as a basis for:    [who advances / interview invite / offer]
  Human review:           [who, what else they see, can they override]

Job relatedness:
  Criteria measured:      [list]
  Tied to duties by:      [job analysis, date, author]
  Validation evidence:    [study type, date, source]

Adverse impact check (29 CFR 1607.4(D)):
  Period and volume:      [dates]  [n applicants]
  Data source:            [voluntary self-ID, stored separately]
  Rates by group:         [group: n, passed, rate, ratio to highest]
  Any ratio under 80%?    [yes/no]   Significance tested? [by whom]

If yes:
  Alternatives considered: [e.g. lower cutoff, remove proxy input,
                            structured human screen]
  Impact of alternatives: [ ]
  Decision and reason:    [ ]

Accommodation route:      [how candidates ask, who handles it]
Vendor version reviewed:  [model/version]   Next review: [date]

Lowering the risk without dropping the tools

• Start from a job analysis. Most disparate impact defenses fail because nobody wrote down why a criterion matters for this job. Do that before configuring any screen.
• Prefer structured human steps where validity is similar. The same questions, asked of everyone and scored against written anchors, are easier to defend than an opaque score. How to reduce bias in interviews covers the method.
• Remove obvious proxies. Graduation years, addresses, gap penalties and school prestige lists are the first things to switch off.
• Keep people deciding. A named reviewer who sees the evidence, not only the score, and who overrides it when it is wrong. Interview Signal is built that way: scores carry the candidate's quoted words from the transcript, and a person decides.
• Check transcript and speech accuracy across speakers if any rating depends on what a tool heard.
• Re-run the numbers on every material change: a new model version, a new cutoff, a new job family.
• Keep the records. Federal rules require keeping application records for at least a year (29 C.F.R. § 1602.14), and some states require longer for automated decision data. How long to keep interview notes has the detail.