Is it illegal for an AI screening tool to penalize résumé gaps?

Penalizing employment gaps is not per se unlawful, but it is a classic disparate impact fact pattern. Career interruptions correlate strongly with pregnancy, childbirth, and caregiving, which are protected under Title VII's sex discrimination provisions as interpreted through the Pregnancy Discrimination Act, and with disability-related leave under the ADA. If a gap penalty produces a materially lower selection rate for women or for applicants who took family leave, the employer must show the practice is job related and consistent with business necessity — and a vendor's claim that the feature 'just predicts performance' is not that showing.

Does the Pregnant Workers Fairness Act apply to algorithmic scheduling?

The PWFA requires covered employers to provide reasonable accommodations for known limitations related to pregnancy, childbirth, or related medical conditions absent undue hardship, and it requires an interactive process. An automated scheduling or shift-bidding system that assigns hours by an availability or reliability score can effectively deny an accommodation before any human conversation happens. The obligation sits with the employer, not the software, and 'the system generated it' is not a defense to a failure-to-accommodate claim.

Is caregiver status a protected class?

There is no standalone federal caregiver protected class, but caregiver discrimination is routinely actionable as sex discrimination under Title VII when it rests on assumptions about who does caregiving, as associational disability discrimination under the ADA when the person cared for has a disability, and as interference or retaliation under the FMLA. A growing number of states and municipalities also protect familial or caregiver status directly. An algorithm that systematically downgrades caregivers reaches all of these routes at once.

How would a plaintiff prove an AI tool discriminated on pregnancy?

Through outcome data rather than intent. Selection rates by sex, by leave history, and by part-time or flexible-schedule request status are computable from the employer's own applicant tracking and HRIS records, and disparate impact requires no proof of discriminatory motive. Discovery typically targets the feature list, model documentation, and any validation study, and the absence of an impact analysis is itself damaging because it establishes that a less discriminatory alternative was never considered.

What should employers audit for pregnancy and caregiver bias in AI hiring tools?

Audit for proxy features first: employment gap length, continuous tenure, availability windows, willingness to work nights or weekends, commute distance, response latency, and any 'reliability' composite. Then compute selection rates by sex and by leave or accommodation history at each funnel stage, not just at final hire. Require the vendor to disclose feature inputs and any validation evidence in writing, document the review, and preserve the ability to override an automated ranking through a human accommodation process.

The Gap in the Résumé Is a Protected Class: AI Screening, Pregnancy, and Caregiver Status

How a Neutral Feature Becomes a Protected Characteristic

Screening models are trained on the historical record of who got hired and who succeeded. In that record, uninterrupted employment histories outperform interrupted ones, because the people with uninterrupted histories were the people the previous system already preferred. The model learns the pattern and encodes it as a feature that looks entirely mechanical: months of continuous employment, or recency of last role, or number of transitions.

The problem is who has gaps. Pregnancy and childbirth produce them. Parental leave produces them. Caring for an aging parent or a child with a disability produces them. Each of those categories is protected — through Title VII as amended by the Pregnancy Discrimination Act, through the ADA's associational provisions, through the FMLA, and in a growing number of jurisdictions through explicit familial-status protections.

Disparate impact doctrine was built for exactly this. A facially neutral practice that falls more harshly on a protected group is unlawful unless the employer proves it is job related and consistent with business necessity, and even then a plaintiff can prevail by showing a less discriminatory alternative was available. Predictive accuracy is not the same as business necessity, and "the model found it useful" is the weakest possible version of that defense.

The Four Proxy Families Worth Auditing First

Continuity and gap features

Employment gap length, months since last role, number of job transitions, tenure averages. The most direct proxy for parental and medical leave, and usually the easiest to find in a feature list.

Availability and schedule breadth

Willingness to work nights, weekends, on-call rotations, or open availability. Caregivers systematically score lower, and the constraint is often the exact thing an accommodation would address.

Composite reliability scores

Vendor-built indices blending attendance, shift-swap frequency, late clock-ins, and callout history. Pregnancy-related appointments and childcare disruption feed straight into these.

Engagement and responsiveness signals

Application response latency, time-to-complete assessments, off-hours activity. These measure free time, and free time is unevenly distributed by caregiving load.

Scheduling Is Where the Accommodation Duty Breaks

Hiring gets the attention, but the sharper exposure sits downstream in workforce management. The Pregnant Workers Fairness Act obliges covered employers to accommodate known limitations arising from pregnancy, childbirth, or related conditions, and to engage in an interactive process to find one. The statute assumes a conversation.

Algorithmic scheduling removes the conversation. A worker who needs to avoid a double shift, sit rather than stand, or hold a fixed morning slot registers in the system as constrained availability. Constrained availability lowers the score. A lower score means fewer hours, worse shifts, or removal from the eligible pool — and that outcome arrives as a schedule, not as a denial anyone made. The employee experiences an adverse action; the employer has no record of having decided anything.

Why this is worse than a human denial: a manager who refuses an accommodation creates a documented decision that can be reviewed and reversed. An automated hours reduction creates a pattern that nobody reviews, applies to everyone in the same situation simultaneously, and produces a clean class of similarly situated plaintiffs with the employer's own scheduling data as the exhibit.

The same structural point applies to algorithmic scheduling discrimination generally and to automated accommodation denials under the ADA, which follow an almost identical mechanism.

Where the Liability Lands

Title VII / PDA. Sex discrimination covering pregnancy, childbirth, and related conditions. Reaches both disparate treatment and disparate impact, and stereotypes about who will be reliable after a birth are the classic treatment theory.
PWFA. Accommodation and interactive-process duties for known pregnancy-related limitations. Automated denial is still denial.
ADA associational discrimination. Adverse action because of an applicant's relationship to a person with a disability — directly implicated by caregiver-availability penalties.
FMLA interference and retaliation. If leave history feeds a score that reduces future opportunity, the leave has been penalized.
State and local law. Several jurisdictions protect familial status, caregiver status, or predictable-scheduling rights outright, and state agencies have been more aggressive on automated decision systems than federal enforcement.

A Concrete Audit Sequence

1. Get the feature list in writing

Ask every screening, ranking, and scheduling vendor for the input variables and any derived composites. A refusal is itself a finding — you cannot defend business necessity for inputs you were never shown, and the refusal belongs in your risk register.

2. Compute selection rates at every stage

Not just final hire. Measure pass rates by sex at résumé screen, assessment, interview invite, and offer. Impact concentrates at a single gate far more often than it spreads evenly, and stage-level data is what tells you which control to change.

3. Segment by leave and accommodation history

For internal tools — scheduling, promotion, performance ranking — compare outcomes for employees who took parental or medical leave in the prior 24 months against those who did not. This is the analysis nobody runs and the one discovery will ask for.

4. Test the counterfactual on gap penalties

Re-rank a historical applicant pool with continuity features removed. If ranking is largely stable, the feature was never carrying the business necessity you would need to defend it. If ranking shifts sharply, you have quantified your exposure.

5. Keep a human accommodation path that outranks the model

Any automated schedule or score must be overridable through a documented accommodation process, and the override must not itself feed back into the reliability score. Systems that penalize the accommodation they just granted are the most common design defect here.

Frequently Asked Questions

The vendor says the model never sees pregnancy or family status. Doesn't that resolve it?

No. Disparate impact analysis does not ask what the model saw; it asks what the model did. Removing the protected attribute while retaining features that correlate with it changes nothing about the outcome distribution, and it removes your ability to measure the disparity. Blindness is not a defense — it is usually the reason nobody noticed.

We only use AI to rank, and a recruiter makes the final call. Are we insulated?

Only if the human review is real. Where recruiters overwhelmingly select from the top of an automated ranking, the ranking is the decision and courts have shown little patience for nominal review. If you rely on human oversight as a control, you need evidence of it: override rates, documented reasons, and candidates advanced from below the algorithmic cutoff.

Does a bias audit under laws like NYC Local Law 144 cover this?

Not by itself. Those audits typically report impact ratios by sex, race, and ethnicity as reflected in reported categories. Pregnancy status, leave history, and caregiver status generally are not among the reported dimensions, so a tool can pass a published audit and still carry the exposure described here. Treat the statutory audit as a floor, not the analysis.

Can we ask candidates about availability at all?

Yes — availability for genuine, essential job requirements is legitimate to assess. The risk arises when availability breadth is used as a general quality signal rather than as a tested requirement, when it is weighted continuously so that broader is always better, and when it is collected before any accommodation conversation can occur.

What is the single highest-value change for a small employer?

Remove employment-gap and continuity features from screening, then compare the resulting rankings to the old ones. It is a one-afternoon change, it eliminates the most direct proxy in the stack, and in most pools it barely moves who reaches the interview list — which tells you what the feature was actually contributing.

The Measurement Is the Whole Defense

Nothing in this area turns on whether an employer meant to disadvantage pregnant applicants or caregivers. It turns on selection rates, on whether anyone looked, and on whether a less discriminatory alternative was available and ignored.

Employers who run the stage-level analysis and act on it have a defensible record even when they find a gap. Employers who never ran it have the gap and no record — and the plaintiff computes the number from their data instead.