Three different things get called “the EEOC’s AI guidance”: a statute from 1964 as amended in 1991, a set of selection-procedure regulations from 1978, and a technical assistance document from May 2023. They have very different force, and only one of them can be withdrawn by an agency changing its mind.
The statute underneath the guidance
Disparate impact liability under Title VII is statutory. It was judge-made in Griggs v. Duke Power Co., 401 U.S. 424 (1971), and Congress wrote it into the text in the Civil Rights Act of 1991, now at 42 U.S.C. § 2000e-2(k). The structure is a three-step burden shift, and every argument about hiring models is an argument about one of the three steps.
- The complaining party identifies a particular employment practice that causes a disparate impact on the basis of race, colour, religion, sex or national origin, and demonstrates that impact.
- The employer must then demonstrate the practice is job related for the position in question and consistent with business necessity.
- Even then, the complaining party wins if it shows an alternative employment practice with less impact that the employer refuses to adopt.
The first step is where scoring models get interesting. Section 2000e-2(k)(1)(B)(i) says that where the elements of a decision process are not capable of separation for analysis, the process may be analysed as one practice. A resume screener whose features cannot be disentangled is a strong candidate for that treatment — which cuts both ways, because it means the employer defends the whole system rather than one feature at a time.
This page describes provisions and their status; it is not legal advice, and nothing here is a substitute for counsel advising on your own selection process, jurisdiction and facts.
Why a 1978 regulation governs a 2026 model
The Uniform Guidelines on Employee Selection Procedures, 29 C.F.R. Part 1607, were adopted jointly in 1978 by the EEOC, the Department of Labor, the Department of Justice and the Civil Service Commission. They apply to any “selection procedure” used as a basis for an employment decision, and § 1607.16(Q) defines that term broadly enough to reach any measure, combination of measures or procedure used in selection — scored application forms, personality inventories, work samples. Nothing in the definition depends on the procedure being a test administered by a human.
That is the whole reason an instrument written before personal computers governs a transformer-based ranker. The Guidelines never regulated a technology; they regulate the act of using a measure to decide who advances. Section 1607.5 then sets out the validation options — criterion-related, content and construct validity — and § 1607.15 sets out what documentation of a validity study has to contain. If you have ever wondered why industrial-organisational psychologists appear in AI hiring disputes, this is why: the defence at step two of the statute is a validation study in the form Part 1607 describes.
The Uniform Guidelines are published in the eCFR at 29 C.F.R. Part 1607.
The four-fifths rule and what it is not
Section 1607.4(D) is the sentence everyone quotes. Its operative words are that a selection rate for any race, sex or ethnic group which is less than four-fifths 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 EEOC’s May 2023 technical assistance document, Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures, applied that arithmetic directly to algorithmic tools: compute the selection rate for each group, divide by the highest, compare to 0.8.
Three things about it are routinely got wrong.
- It is a rule of thumb, not a safe harbour. The same section says smaller differences may constitute adverse impact where they are significant in both statistical and practical terms, and greater differences may not where the numbers are small. A ratio of 0.83 is not a defence.
- It is about selection rates, not scores. A model that produces a continuous score has no selection rate until somebody sets a threshold. Move the cut-off and the ratio moves. This is why the choice of threshold is itself the employment practice in many disputes.
- Passing it does not validate anything. The four-fifths comparison is a screen for whether step one of the statute is likely to be met. It says nothing about job-relatedness, which is a separate evidentiary burden with its own regulation.
A practical consequence for anyone instrumenting a pipeline: you cannot compute the ratio without knowing applicants’ protected class, which most employers deliberately do not collect at the screening stage. The usual answer is voluntary self-identification held separately from the selection data, which is permitted but produces incomplete denominators. Imputation from names and geography — the technique Colorado’s insurance regulator mandates, discussed in the Colorado Division of Insurance testing rule — has no equivalent blessing in the employment context.
When the vendor’s tool becomes the employer’s liability
The 2023 technical assistance was explicit that an employer may be responsible for a tool it did not build. The theory is agency: where the vendor acts on the employer’s behalf in administering a selection procedure, the employer does not escape Title VII by outsourcing. “The vendor told us it was bias-free” is not a defence recognised anywhere in Part 1607, and § 1607.7 sets conditions on when an employer may rely on a validity study conducted elsewhere — broadly, the jobs and the candidate population have to be comparable, and the study has to meet the same documentation standard.
Whether the vendor itself can be liable is a separate and genuinely unresolved question. It is being litigated in the Workday screening litigation, where the theory is that a screening provider can be an “employment agency” or an agent of the employers using it. Nothing about that theory is settled, and a district court ruling allowing a claim to proceed decides that the claim is plausible, not that it is correct.
What the iTutorGroup settlement decided
EEOC v. iTutorGroup, Inc. was filed in the Eastern District of New York in 2022 and settled in August 2023 for $365,000 under a consent decree. It is widely described as the EEOC’s first settlement involving hiring software, and that description is doing a lot of work.
The conduct alleged was that the company’s application software automatically rejected female applicants aged 55 or over and male applicants aged 60 or over, on the basis of a date of birth entered in the form. That is a hard-coded filter, not a learned model, and the claim was brought under the Age Discrimination in Employment Act rather than Title VII. What the case establishes is that automating a rule does not change its legal character — useful, and much narrower than “the EEOC has taken on AI hiring”.
It also decided nothing at all as precedent. A consent decree is a settlement entered by a court; it contains no finding of liability and binds only the parties. Cited as authority for what the law requires, it proves the agency’s willingness to sue, and no more. The longer treatment is in the dedicated page on the settlement.
The 2025 change in enforcement posture
Two things happened in 2025 that a page written in 2024 would get wrong. Executive Order 14281, “Restoring Equality of Opportunity and Meritocracy”, signed in April 2025, directs federal agencies to deprioritise enforcement of statutes and regulations that impose disparate-impact liability and to evaluate amending or repealing regulations that do so — the Uniform Guidelines are named among the candidates. Separately, the EEOC removed several artificial-intelligence-related technical assistance documents from its website, and the Commission spent much of 2025 without the quorum it needs to issue or rescind formal guidance.
Enforcement posture is the fastest-moving thing on this page. Before relying on any statement in this section, check the current text at the EEOC’s guidance library and search the Federal Register for the executive orders by number at federalregister.gov. This page is not legal advice.
What did not change is the part that matters most to an employer deciding whether to deploy a screener. An executive order directs the executive branch; it does not amend 42 U.S.C. § 2000e-2(k), which Congress enacted and only Congress can repeal. Private plaintiffs bring the overwhelming majority of Title VII cases and are not bound by federal enforcement priorities. State fair-employment agencies operate under their own statutes, several of which now address automated decision tools directly — see Illinois HB 3773 and New York City Local Law 144. And a validity study prepared to Part 1607 standards remains the strongest evidence available at step two of the statute regardless of which agency is currently enforcing it.
The honest summary is that the guidance document’s availability changed, the agency’s priorities changed, and the exposure did not. Anyone treating 2025 as an all-clear is reading an executive order as though it were an amendment to the Civil Rights Act.
Top comments (0)