Most conversations about AI hiring bias are moral ones. This page isn't. The US has had a settled legal framework for discriminatory selection procedures since 1971, it was designed for exactly this — a test that screens people out for reasons nobody can articulate — and it applies to your model whether or not anyone at your company has read it.
Two ways to discriminate
- Disparate treatment
- Intentional discrimination. Requires motive. For AI: feeding the model a protected characteristic, or using the tool as cover for a decision already made.
- Disparate impact
- A facially neutral practice, applied neutrally, that falls more harshly on a protected group. No intent required. This is the theory that reaches algorithmic scoring, and the one that should shape how you build.
Griggs v. Duke Power Co., 401 U.S. 424 (1971) established it. Duke Power required a high-school diploma and a general intelligence test for certain jobs. Neither was shown to relate to the work. The Supreme Court held that Title VII "proscribes not only overt discrimination but also practices that are fair in form, but discriminatory in operation," and — the line worth memorising — that "good intent or absence of discriminatory intent does not redeem employment procedures."
The instrument struck down in 1971 was a general aptitude measure with no demonstrated relationship to the job. A generic ML resume-ranker trained on "who we hired before" is structurally identical: a general predictor, applied to everyone, whose connection to job performance nobody has established. The law didn't need updating for AI. AI walked into it.
Who has to prove what
The Civil Rights Act of 1991 codified the structure at 42 U.S.C. § 2000e-2(k), and the burden allocation is the part that should worry a vendor.
- Plaintiff demonstrates that a particular employment practice causes a disparate impact — § 2000e-2(k)(1)(A)(i).
- Employer must then demonstrate the practice is "job related for the position in question and consistent with business necessity." This is a burden of production and persuasion, not merely of pointing. You must affirmatively produce evidence.
- Plaintiff can still win by showing a less discriminatory alternative existed and the employer "refuses to adopt" it — § 2000e-2(k)(1)(A)(ii).
A plaintiff must normally isolate each particular challenged practice. But the statute continues: if the complaining party "can demonstrate to the court that the elements of a respondent's decisionmaking process are not capable of separation for analysis, the decisionmaking process may be analyzed as one employment practice."
Read that against a black-box model. The intuition that an unexplainable system is hard to attack is exactly backwards: inseparability is the plaintiff's argument, not yours. Where you cannot show which factor drove the outcome, the court is invited to treat your entire pipeline as a single practice — and then hand you the burden of justifying the whole thing at once. Explainability isn't a nice-to-have here. It is what keeps the analysis granular enough to defend.
The four-fifths rule
The Uniform Guidelines on Employee Selection Procedures (UGESP), 29 CFR Part 1607, were adopted jointly in 1978 by the EEOC, DOL, DOJ and OPM. The four-fifths rule lives at 29 CFR 1607.4(D):
"A selection rate for any race, sex, or ethnic group which is less than four-fifths (4/5) (or eighty percent) 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 arithmetic is simple. Selection rate per group = selected ÷ applicants. Impact ratio = that group's rate ÷ the highest group's rate. If 30% of men advance and 15% of women do, the ratio is 0.15 / 0.30 = 0.50 — a clear failure.
Note the hedges in the text itself: "generally" and "by the Federal enforcement agencies". This is an enforcement-screening heuristic, not a statutory liability standard. The regulation immediately qualifies itself in both directions:
"Smaller differences in selection rate may nevertheless constitute adverse impact, where they are significant in both statistical and practical terms…" and "Greater differences in selection rate may not constitute adverse impact where the differences are based on small numbers and are not statistically significant…"
So: passing 4/5ths across a large applicant pool can still be statistically significant adverse impact. Failing it across twelve applicants may be noise. Courts routinely work from standard-deviation analysis alongside or instead of the ratio. Treat 0.80 as a smoke alarm, not a certificate.
Validation under UGESP
"Job related and consistent with business necessity" is not an assertion you make — it is something you validate. Albemarle Paper Co. v. Moody, 422 U.S. 405 (1975) tied the defence to the EEOC guidelines and to validation evidence: the procedure must be "predictive of or significantly correlated with important elements of work behavior which comprise or are relevant to the job."
29 CFR 1607.5(A) recognises three strategies, with technical standards at 1607.14. All of them require a job analysis first (1607.14A).
| Strategy | Cite | What it requires |
|---|---|---|
| Criterion-related | 1607.14B | A statistical relationship between scores and job-performance measures, significant at the 0.05 level, on a sample representative of the relevant labour market. |
| Content | 1607.14C | The procedure is a representative sample of the job's content — observable work behaviours or products. |
| Construct | 1607.14D | Empirical evidence linking a measured trait to critical work behaviours, typically backed by criterion-related evidence. The hardest of the three. |
Content validity — the cheapest strategy to establish — is expressly inappropriate for procedures purporting to measure "traits or constructs, such as intelligence, aptitude, personality, commonsense, judgment, leadership, and spatial ability."
That sentence rules out most of the category. Personality inference, "culture fit", video- or voice-based trait scoring, game-based assessment — these are construct measures. 1607.14C is unavailable to them and 1607.14D is demanding. So when a vendor says "validated", the only useful response is: which strategy, against what job analysis, at what significance level, on what sample? An answer that doesn't contain those four things isn't a claim of validity. It's a claim of having heard the word.
The trap: you can't fix the number
Suppose you measure impact and the ratio comes back 0.62. The obvious engineering instinct — adjust the thresholds by group until it clears 0.80 — is not merely inadvisable. It is separately illegal.
42 U.S.C. § 2000e-2(l) bans race-norming: employers may not "adjust the scores of, use different cutoff scores for, or otherwise alter the results of, employment related tests on the basis of race, color, religion, sex, or national origin." And Ricci v. DeStefano, 557 U.S. 557 (2009) held that discarding test results to avoid disparate-impact liability was itself disparate treatment, absent a strong basis in evidence.
You are required to look for adverse impact. You are forbidden from fixing it at the output by group. The only lawful path left runs backwards through the system: find the impact before deployment, then remediate by redesign and validation — different criteria, better job analysis, a less discriminatory alternative that does the same job.
Which means bias testing has to be part of building the assessment, not a report you run afterwards on a decision you have already made. A platform that only tells you your ratio at the end has handed you a problem you are not allowed to solve.
What changed in 2025 — and what didn't
Federal enforcement posture shifted in 2025. Federal law did not. The distinction is the most commonly muddled point in this area, and getting it wrong in either direction is expensive.
What changed
- The EEOC's 2022 ADA/algorithms guidance and 2023 Title VII AI guidance were removed in January 2025. Both canonical URLs now return 404 — we checked.
- Executive Order 14281, "Restoring Equality of Opportunity and Meritocracy" (23 April 2025), states a policy "to eliminate the use of disparate-impact liability in all contexts to the maximum degree possible," and directs agencies to "deprioritize enforcement of all statutes and regulations to the extent they include disparate-impact liability, including but not limited to 42 U.S.C. 2000e-2."
What did not change
- An executive order cannot repeal a statute. § 2000e-2(k) was enacted by Congress and is repealable only by Congress or narrowable by the courts.
- Private plaintiffs kept their cause of action. Title VII suits don't depend on EEOC enthusiasm — a charge must be filed for exhaustion, but on a decline to sue the agency issues a right-to-sue letter and the plaintiff proceeds. Reduced federal enforcement plausibly shifts risk toward private litigation, not away from it.
- EO 14281 is self-limiting. Its own general provisions state it "is not intended to, and does not, create any right or benefit, substantive or procedural, enforceable at law" — so it is not a defence an employer can plead.
- UGESP was not revoked. EO 14281's revocations reach Title VI regulations, not 29 CFR Part 1607. Repealing UGESP would require APA notice-and-comment rulemaking.
- State and local law is untouched, and courts are not bound by executive policy.
State & local law in force
A great deal of published material still says the Colorado AI Act (SB 24-205) takes effect on 1 February 2026, or that it was delayed to 30 June 2026. Both are now wrong. SB 26-189, signed 14 May 2026, repealed and reenacted it with a new effective date of 1 January 2027 — and the duty of reasonable care to avoid algorithmic discrimination was dropped in favour of a disclosure/transparency framework with human review and correction rights. Per the legislature's summary, it "does not create a new private right of action but establishes how fault is allocated between developers and deployers."
| Jurisdiction | Status | What it does |
|---|---|---|
| NYC Local Law 144 |
In force since 5 Jul 2023 |
Annual bias audit by an independent auditor (6 RCNY § 5-301(a)); publish a summary of results including impact ratios by sex, race/ethnicity and intersectional categories (§ 5-303(a)); 10 business days' notice to NYC candidates. Note it mandates publication of the ratio, not achievement of 0.80. |
| Illinois HB 3773 |
In force since 1 Jan 2026 |
Amends the Human Rights Act. Civil rights violation to use AI with the effect of discriminating — unintentional discrimination is covered. Express ban on zip code as a proxy. Notice required when AI is used. |
| California FEHA / CRC ADS regs |
In force since 1 Oct 2025 |
Covers any system that "makes a decision or facilitates human decision making" — so "a human reviews it" is not an exemption. Agent liability extended to vendors. Record retention extended to four years, expressly covering ADS data. Bias testing isn't mandated, but its presence — or absence — is relevant evidence. |
| Colorado SB 26-189 |
Not yet 1 Jan 2027 |
Repealed and replaced SB 24-205. ADMT disclosure regime; duty of care dropped. No AI-discrimination duty of care is currently in force in Colorado. |
Mobley v. Workday
Mobley v. Workday, Inc., No. 3:23-cv-00770 (N.D. Cal.) is the case that made this a vendor problem rather than only an employer problem.
- July 2024 — Judge Rita Lin held that an AI vendor may be directly liable under Title VII, the ADA and the ADEA on an agent theory: Workday's tools screen, score, rank and recommend, and so perform a function traditionally exercised by the employer.
- 16 May 2025 — preliminary collective certification under the ADEA, nationwide, covering applicants aged 40+ denied employment recommendations via the platform since 24 September 2020.
- 22 June 2026 — a further motion to dismiss granted in part and denied in part. Title VII and ADEA disparate-impact claims, the ADA claim, an ADA proxy-discrimination theory and California FEHA claims survived; the direct-employer theory was dismissed. The case is live and in discovery.
Why it matters beyond Workday: a single vendor's model deployed across thousands of employers turns a per-employer dispute into one collective action spanning the customer base — the shared screening logic is precisely the "common policy" that supports certification. And California's FEHA regulations now give the agent theory a statutory-regulatory anchor in the largest state economy.
Records you must keep
29 CFR 1602.14 requires personnel and employment records — expressly including "application forms submitted by applicants and other records having to do with hiring" — to be preserved for one year from the making of the record or the personnel action, whichever is later.
The second limb matters more. Once a charge is filed, the employer "shall preserve all personnel records relevant to the charge or action until final disposition" — and relevant records expressly extend beyond the complainant to employees in similar positions and to application forms and test papers of unsuccessful applicants for the same position.
Two consequences worth sitting with. First, the retention clock is not knowable in advance: a routine one-year purge that fires after a charge lands destroys evidence you were obliged to keep. Second, a litigation hold reaches the comparator pool — model inputs, scores and rankings for people who never complained.
Separately, 29 CFR 1607.4(A) requires users to maintain records disclosing the impact of their selection procedures by race, sex and ethnic group. That is an affirmative, continuous duty — it does not wait for a charge. Note what it presupposes: you cannot compute an impact ratio retrospectively if you never recorded the selection events. More on retention and holds.
Where HireBeep fits
Every US obligation on this page lands on the employer. No vendor can absorb them. What a platform can do is make them satisfiable rather than theoretical — which is what the design is for.
- One consistent rubric, bias-reviewed before publish — impact found at design time, when redesign is still the available remedy and score adjustment isn't.
- Adverse-impact monitoring — the 1607.4(A) records exist because the selection events were recorded as they happened.
- Explainable grading — a written rationale keeps the analysis separable, so § 2000e-2(k)(1)(B)(i) doesn't collapse your process into a single unjustifiable practice.
- Exportable audit trails — a hold can reach the comparator pool because the comparator pool was logged.
This is a sourced explainer, not advice, and it is not a substitute for counsel who knows your jurisdictions and your data. Where it matters, we have cited the primary law so you can check us.
- 42 U.S.C. § 2000e-2 (Title VII, unlawful employment practices) — Cornell LII
- 29 CFR 1607.4 (UGESP — records, four-fifths rule) — Cornell LII
- 29 CFR 1607.5 (validation strategies) — Cornell LII
- 29 CFR 1607.14 (technical standards) — Cornell LII
- 29 CFR 1602.14 (record retention) — GovInfo
- Griggs v. Duke Power Co., 401 U.S. 424 (1971); Albemarle Paper Co. v. Moody, 422 U.S. 405 (1975); Ricci v. DeStefano, 557 U.S. 557 (2009)
- Executive Order 14281, "Restoring Equality of Opportunity and Meritocracy", 23 April 2025 — whitehouse.gov
- Cooley, "Gone but not forgotten: federal laws still apply despite guidance disappearance act" — cooley.com
- NYC DCWP, adopted rules for Automated Employment Decision Tools (6 RCNY subch. T) — rule text (PDF)
- Colorado SB 26-189 (repeal & reenactment of the Colorado AI Act) — leg.colorado.gov; Seyfarth — analysis
- California Civil Rights Department, ADS regulations final text — PDF
- Seyfarth on Mobley v. Workday agent theory — analysis; Duane Morris on the June 2026 ruling — analysis