Policy
An AI Requirement Answers to the ADEA, Not to the Four-Fifths Rule
An AI-fluency requirement that leaves nobody over fifty on the shortlist can be an age discrimination problem, and it runs under the Age Discrimination in Employment Act rather than Title VII. The ADEA reaches applicants forty and over, disparate-impact claims are available under it, and the employer carries the burden of showing the practice rests on a reasonable factor other than age. Compute selection rates by age band at each stage, then name the job tasks the requirement stands in for.
The takeThe requirement most teams actually wrote is a tenure test wearing a skills label. Asking for three years of hands-on AI work in 2026 selects for people whose employer bought a licence early and who changed jobs recently, which is a fact about employers more than about candidates. The defensible version of the same line is duller and shorter: name what a person has to be able to do on a task, and let the candidate show it with whatever tool they happened to have.
Where Olive fits
Open a role and see what the work shows
On an age-impact question, a number a tool assigned is not a job task anybody can point at. Olive produces no composite, and no decision about a candidate is made by a machine alone: a person writes or edits each of the six findings, every one carrying the timestamped excerpt it rests on, and every released report exports with its rubric, scorer and bank versions attached.
Assess your shortlistWhich law covers a shortlist that thins by age?
The Age Discrimination in Employment Act covers it, and the test under that statute is not the one most hiring teams have run. The ADEA's prohibitions are limited to individuals who are at least forty years of age 1. A claim can rest on the effect of a neutral requirement alone, so a line nobody wrote with age in mind is still reachable.
The Supreme Court settled the impact question in March 2005. Smith v. City of Jackson holds that the ADEA authorizes recovery in disparate-impact cases comparable to Griggs, and says in the same syllabus that the theory's scope is narrower under the ADEA than under Title VII 2.
The narrowing sits in one clause.
Section 4(f)(1) permits an otherwise prohibited action where the differentiation is based on reasonable factors other than age 2. An employer defending an AI requirement is not asked for business necessity. The employer is asked for a reasonable factor.
A reasonable factor is a lighter burden than business necessity, and it is still the employer's to carry. The EEOC's regulation is explicit: whenever the reasonable-factors-other-than-age defense is raised, the employer bears the burdens of production and persuasion to demonstrate the defense 3.
The arithmetic most teams already run does not reach any of this.
The four-fifths paragraph in the Uniform Guidelines speaks of a selection rate for any race, sex or ethnic group falling below eighty percent of the rate for the highest group 4. Age appears nowhere in the paragraph. A team that has run the four-fifths math and found nothing has answered a different question.
Why a years-of-exposure line reads on job changes
AI tools reached ordinary desks only a few years ago, and exposure at work can depend on what an employer bought and when a person last started somewhere new. A requirement written as years of hands-on AI experience does not ask what someone can do with a model. It asks where they have been sitting since the tools arrived, and job tenure has a steep gradient by age.
In January 2024 the median US wage and salary worker had been with their current employer 3.9 years, the lowest reading since 2002 5.
The figure splits hard by age. Median tenure ran 9.6 years for workers aged 55 to 64, against 2.7 years for workers aged 25 to 34, in a January snapshot of job spells that were still running 5.
If the median worker in the older band had held the same job since roughly 2014, the years in which generative tools reached ordinary desks fell inside a job they already had, under whatever licence their employer chose to buy.
The younger figure carries no message about drive. A worker in the younger band has had fewer years in which to accumulate tenure at all.
What the requirement selects on, then, is a purchasing decision at a company the candidate may no longer work for.
The requirement is a proxy, a line standing in for something else, and proxies survive deletion because the record rebuilds them. Cutting the number out of the posting does not settle it either, which is why what to require instead of years of AI experience is the harder half of the job.
Run the selection rates by age band, stage by stage
Compute the advance rate at every stage where people leave, split at forty and again in ten-year bands above it, and hold the denominator to the people actually considered at that stage. The stage matters more than the total, because a requirement that thins the pool in the posting is invisible in a shortlist ratio.
A knockout question in the application form, a recruiter's read of a resume and a scored exercise produce three different rates. Only one of them may be doing the work.
Then read the ratio for what it is. Eighty percent is a rule of thumb the agencies wrote to decide where to look. The same paragraph says a smaller gap can still be adverse impact where it is statistically and practically significant, and a larger one may not be where the numbers are small and not significant 4.
Under the ADEA a claimant carries a step of this too. Smith requires the specific test, requirement or practice producing the adverse impact to be identified 2.
An AI-fluency line is exactly that kind of specific practice, and it is already written down in the posting.
Keep the record at the level of the person. A rate can be recomputed from individual records later, and individual records cannot be recovered from a rate. What a position statement has to show about the AI round is written out of those records or out of nothing.
What the reasonable-factor defense asks an employer to show
The defense asks for two things, sound design and sound administration, and the second is where AI requirements come apart. The regulation says a reasonable factor other than age is one the employer shows was both reasonably designed to further or achieve a legitimate business purpose and administered in a way that reasonably achieves that purpose 3. A requirement can be defensible as written and unreasonable in how a recruiter applied it.
The regulation then lists what counts, and the first item is the extent to which the factor is related to the employer's stated business purpose 3. The fourth item is the extent to which the employer assessed the adverse impact of the practice on older workers 3.
So the selection-rate exercise above also serves the defense. Under that list, having run it is part of the answer.
No single consideration decides the question, by the regulation's own terms 3. Confirm the current text, and any state law that reaches further, with employment counsel; none of this is legal advice.
The repair is to state the business purpose in job tasks. Reviewing a model's draft against the source document and rejecting it on stated grounds is a task a person either does or does not do. Three years using AI tools is a purchasing history.
Then rewrite the posting so it accepts the behavior wherever it was learned. The or-equivalent line is where that repair lands, and what makes an AI requirement in a posting defensible sets out the general form the wording has to take.
Should the shortlist be reopened, or the requirement rewritten?
Rewrite the requirement first, and let that decide what happens to the shortlist. A list repaired by adding people back, without touching the rule that produced it, produces the same list next quarter. Re-read the line, decide whether it names a task or a tenure, and only then decide who was removed by something nobody can defend.
A short list settles almost nothing. Nobody over fifty on a shortlist of a handful is a number with very little in it, so the case for the rewrite rests on what the line asks for, not on the count.
The enforcement record is also thinner than the commentary around it. In EEOC v. iTutorGroup three tutoring companies agreed to pay $365,000 to settle claims that they programmed their tutor application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older 7.
A hard-coded date-of-birth cutoff is a rule somebody typed, not a model that learned anything, and the case ended in a settlement 7.
The case to watch is Mobley v. Workday. In May 2025 the court granted preliminary collective certification of an ADEA claim, and in July 2025 held that the collective includes applicants scored, ranked or screened by the vendor's AI features 6.
Nothing in that docket is a finding that anyone discriminated. Preliminary certification permits notice and nothing more, decertification remains available, and the case was still moving in 2026 6.
The shortlist on the desk is now a document about the requirement.
Read the line again, write down the task it was standing in for, and put the rewritten version in the posting before the next req opens.
Common questions
Does the ADEA cover disparate impact, or only intentional discrimination?
Both, and the impact route has been available since 2005. In Smith v. City of Jackson the Supreme Court held the ADEA authorizes recovery in disparate-impact cases comparable to Griggs, while saying in the same syllabus that the theory's scope is narrower under the ADEA than under Title VII. The narrowing is the reasonable-factors-other-than-age clause in section 4(f)(1), which permits an otherwise prohibited action where the differentiation is based on reasonable factors other than age. A claimant still has to identify the specific requirement or practice producing the disparity. Check the current state of the law with employment counsel.
Does the four-fifths rule apply to age?
No. The four-fifths paragraph sits inside the 1978 Uniform Guidelines, and its own text speaks of a selection rate for any race, sex or ethnic group falling below eighty percent of the rate for the highest group. Age is not in it. The ratio is still a useful first look at a set of numbers split at forty, as long as nobody treats eighty percent as a line that clears an employer. The same paragraph says a smaller gap can still be adverse impact where it is statistically and practically significant, and that a larger one may not be where the numbers are small and not significant.
Is a reasonable factor other than age easier to show than business necessity?
Lighter, and still the employer's to prove. The EEOC's regulation says that whenever the reasonable-factors-other-than-age defense is raised, the employer bears the burdens of production and persuasion. What has to be shown is that the practice was reasonably designed to further or achieve a legitimate business purpose and administered in a way that reasonably achieves that purpose. Design alone will not carry it: a requirement written well and applied loosely fails the second half. The regulation adds that no single consideration on its list decides the question.
Can an AI requirement go in a posting at all?
Yes. What creates the exposure is the shape of the ask. A requirement stated as years of on-the-job exposure to a tool that is a few years old reads on when a candidate last changed employers and on what that employer licensed. A requirement stated as an observable act on a task, such as checking a model's claim against a source before it ships, can be met by someone who learned it anywhere, including outside paid work. Write the second kind, accept equivalents, and keep the record of how the line was applied.
What age bands should the selection rates use?
Split at forty, where the ADEA's coverage begins, then in ten-year bands above it so a gradient stays visible. Compute the rate at each stage separately: application to screen, screen to interview, interview to shortlist. The denominator is the people considered at that stage, not everyone who applied. Small samples settle nothing, and a shortlist of a handful carries almost no information, so the exercise finds which stage to look at.
What if the AI requirement really is necessary for the job?
Then say what the job needs in the terms the job uses. A role where a person reviews model output before it reaches a client has a real requirement, and the requirement is about reviewing output rather than about elapsed months with a product. Write the task, write how it gets assessed, and apply the same line to everyone who reaches that stage. Two of the reasonable-factor considerations point at exactly that: how far the factor is related to the stated business purpose, and whether the employer assessed the impact on older workers.
References
- 1. 29 U.S.C. 631 - Age limits (Age Discrimination in Employment Act) ✓ uscode.house.gov Supports the claim that the ADEA reaches applicants forty and over. Subsection (a) reads: "The prohibitions in this chapter shall be limited to individuals who are at least 40 years of age."
- 2. Smith v. City of Jackson, Mississippi (No. 03-1160), syllabus ✓ law.cornell.edu Decided March 30, 2005. Supports three claims: that "the ADEA authorizes recovery in disparate-impact cases comparable to Griggs"; that "the disparate-impact theory's scope is narrower under the ADEA than under Title VII" because section 4(f)(1) permits an otherwise prohibited action "where the differentiation is based on reasonable factors other than age"; and that a claimant must identify a specific test, requirement or practice rather than point at disparate outcomes.
- 3. 29 CFR 1625.7 - Differentiations based on reasonable factors other than age ✓ law.cornell.edu Source credit [46 FR 47726, Sept. 29, 1981, as amended at 77 FR 19095, Mar. 30, 2012]. Supports the burden statement at (d), that whenever the reasonable-factors-other-than-age defense is raised the employer bears the burdens of production and persuasion; the two-part test at (e)(1), reasonably designed to further or achieve a legitimate business purpose and administered in a way that reasonably achieves that purpose; the first and fourth considerations at (e)(2)(i) and (e)(2)(iv), relatedness to the stated business purpose and the extent to which the employer assessed the adverse impact on older workers; and (e)(3), that no specific consideration or combination of considerations is dispositive.
- 4. 29 CFR 1607.4 - Information on impact (Uniform Guidelines on Employee Selection Procedures, 1978) ✓ law.cornell.edu Supports the claim that the four-fifths paragraph names race, sex and ethnic group and not age, and that the same paragraph treats eighty percent as an enforcement rule of thumb rather than a threshold: smaller differences may still constitute adverse impact, and the rule is about what federal enforcement agencies will generally regard as evidence.
- 5. Employee Tenure Summary (Employee Tenure in 2024, USDL-24-1971) ✓ web.archive.org Supports the January 2024 median tenure of 3.9 years, the lowest since January 2002, and the age split of 9.6 years for workers aged 55 to 64 against 2.7 years for workers aged 25 to 34. Used here as a snapshot of job spells still running rather than a measure of how long a job lasts. Confirmed against the fact-pool quote fetched 2026-08-26; both the live BLS page and the Wayback snapshot refuse automated requests.
- 6. Mobley v. Workday, Inc., 3:23-cv-00770 (N.D. Cal.) ✓ clearinghouse.net Supports the May 16, 2025 grant of preliminary collective certification on the ADEA claim, the July 2025 holding that the collective includes applicants scored, ranked or screened by Workday's HiredScore AI features, and the article's caveats: no merits ruling, decertification still available, docket still moving in 2026.
- 7. iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit ✓ eeoc.gov Supports the $365,000 settlement and the EEOC's description that the companies "programmed their tutor application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older," which the article uses as a hard-coded cutoff rather than as a model. The release states the money goes to applicants the software rejected and requires hiring training, a new anti-discrimination policy, injunctions against age- and sex-based rejection and against requesting birth dates, with EEOC monitoring for at least five years. It carries no liability language and does not use the phrase consent decree.
7 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance, not legal advice. Hiring rules differ by state and country and change often; check anything here against your own counsel before you act on it.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.