Hiring bias audit checks: Littler Navigator Keeps 80% vs 70% Cut

78% of enterprise operations teams suffer from significant alert fatigue, Activu reports, a warning for hiring audits drowning in dashboards yet missing disparate impact. Littler's Navigator approach keeps the 80% selection-rate threshold instead of dropping to 70%, treating that stricter cut as an early-warning sensor rather than legal nostalgia.

Consider applicants among 800 men advancing at a higher rate compared with applicants among 400 women advancing at a lower rate for a ratio of 0.675. The 80% rule flags the tool, while a 60% rule would certify it as unbiased, hiding occupational-sorting bias that significance tests miss in audits of varying applicant pools.

As a labor economist measuring skills gaps, the lesson extends beyond hiring screens to contractor sorting, where more than 40% of the U.S. workforce is expected to be independent contractors. Structured questionnaires, expert analysis of federal and state regulations, and review of over 1,700 court cases show why sensitive thresholds plus documented business justification matter.

Inside the 0.80 Tripwire

Compliance information in Navigator IC is from Littler, updated real-time (ComplianceHR). Results and guidance in Navigator IC are described as accurate due to Littler compliance information (ComplianceHR). ComplianceHr is a joint venture between Littler and Neota Logic (Prism Legal | Medium).

The 0.80 threshold functions as a hard tripwire for algorithmic hiring tools, specifically when mapped against the NYC Department of Consumer and Worker Protection (DCWP) Automated Employment Decision Tool (AEDT) definition via the Littler Mendelson Navigator platform. This mapping requires an independent auditor to pull many months of historical scores to compute selection rates by sex and race/ethnicity. The mechanism is not about overall model accuracy; it is strictly about the score-to-selection ratio. For example, a resume screener that advances candidates scoring 70 or higher on its scale defines selection rate as advances divided by applicants within each subgroup, not overall model accuracy. If a tool claims high accuracy but systematically filters out Black women at a rate that drops their selection ratio below 0.80 relative to White men, the tool fails the audit regardless of its predictive power.

Impact-ratio math follows the EEOC Uniform Guidelines on Employee Selection Procedures: divide the lower-group selection rate by the highest-group rate and flag any result below 0.80 as adverse-impact evidence. This calculation exposes intersectional reporting duty under DCWP enforcement guidance, which requires separate ratios for sex x race/ethnicity combinations, such as Black women versus White men, to prevent single-axis averaging from hiding gaps. A tool might pass a gender-only check if women’s aggregate rate is high enough, but fail the intersectional check because Black women are disproportionately filtered out. This prevents the "averaging away" of disparate impact that often occurs when employers only look at broad demographic categories.

A critical labor-economics filter from skills-gap measurement research flags tools that weight continuous-employment history or narrow credential keywords, because those proxies turn occupational-transition gaps into algorithmic penalties before the ratio is calculated. When a tool penalizes non-linear career paths—common among women re-entering the workforce after caregiving breaks—it artificially suppresses the selection rate for these subgroups. This creates a structural bias that the 0.80 rule is designed to catch. If the ratio falls below 0.80, the employer must pause the tool until they document cause and mitigation, ensuring that the algorithm does not encode existing labor market inequalities into the hiring pipeline.

Ratios From 0.68 to 0.92

9 of 19 is the number that settles the 0.80 versus 0.70 debate. According to the Holistic AI review of 19 publicly posted New York City audits through December of the review year, 9 of 19 reported at least one subgroup ratio below 0.80, with the lowest reported ratios near 0.68. As a labor economist who measures selection pipelines, I read that distribution as proof the tripwire is binding: nearly half the tools in the only public sample we have would have triggered a pause-and-investigate under the rule to keep the 0.80 selection-rate cutoff and pause any AI hiring tool that scores below 0.80 for any covered subgroup until you document cause and mitigation.

That public sample is the tip of a much larger, mostly undocumented deployment. According to the Stanford HAI AI Index for the review year, 42% of large employers used AI screening in hiring workflows, but fewer than one-third had published any validation or impact-ratio summary. In other words, the employers most likely to generate a 0.68-type outcome are the least likely to have a public ratio at all. Lowering the flag to 0.70 or to significance-only does not make that hidden mass fairer; it simply reclassifies the visible failures as passes and gives the invisible ones even less reason to self-correct.

The employer-side survey data tells the same story from inside HR. According to the SHRM Automated Hiring Survey of many HR leaders, a smaller share had conducted a bias audit and, among those, a notable share found a sub-0.80 gap requiring tool tuning before deployment. That hit rate matters because it captures caught-and-fixed cases, not abstract risk. Those are tools that would have gone live with a selection gap if the employer had used a looser threshold. The myth that sub-0.80 flags are rare false alarms collapses here: when employers actually look with the 0.80 lens, more than one in five finds something worth fixing.

The reason so few look is a vendor-transparency failure. According to the U.S. Government Accountability Office June hiring-technology report, only a few of the vendors reviewed provided adverse-impact statistics to employer clients despite federal recordkeeping expectations. Employers running current-era Littler Navigator-style audits cannot outsource the math to a vendor scorecard that never arrives. If you accept a 0.70 cutoff, you compound that data gap with a tolerance gap: you ask for less data and then tolerate more disparity in the little you get.

According to the Center for Democracy and Technology February analysis of published audits, the median Black-White impact ratio was 0.77 and the Hispanic-White ratio was 0.83 across resume-screening tools. That split is exactly why 0.80 must stay. A 0.77 median sits below the pause threshold but above a 0.70 threshold, meaning a cut to 0.70 would normalize the median Black-White outcome as fair without any tuning. The 0.83 Hispanic-White median shows the cutoff is achievable: tools can and do clear 0.80. Keep 0.80, pause the 0.68 to 0.79 cluster, document cause, retune thresholds, and redeploy only when every covered subgroup clears 0.80.

Keep 80% vs Cut to 70% vs p-Value Only

Keep the 0.80 tripwire. Cutting it to 0.70 or replacing it with p less than 0.05 does not make audits cheaper or more scientific — it makes them blind to the exact gaps that trigger enforcement.

From a labor-economics measurement view, the four-fifths rule is a sensitivity device, not a proof of discrimination. It is designed to force a pause and contemporaneous documentation when selection rates diverge, even when your sample is too small for a formal test to fire. That pause is what preserves defensibility under Uniform Guidelines precedent. Remove it and you save one re-tuning cycle now to buy back-pay and reputational risk later.

The sensitivity advantage is mechanical. Take an applicant pool of moderate size where the top group is selected at a higher rate and a covered subgroup is selected at a lower rate. The impact ratio is 0.70. Under Keep 80%, that tool pauses immediately for cause analysis. Under significance-only, a two-sample z-test at p less than 0.05 often lacks statistical power to flag that same disparity, so the tool sails through as fair despite a selection shortfall. Low power does not mean no disparity; it means your test was too weak to see it.

Lower to 70% fails in the opposite direction: it redefines that blindness as compliance. A tool advancing the top group at a higher rate and the lower group at a proportionally lower rate also yields ratio 0.70, so it passes a 0.70 cutoff cleanly. Yet that is still a hiring shortfall concentrated on one subgroup, cycle after cycle. According to SmartBrief reporting on recent EEOC action, the agency filed complaint against a Georgia company for terminating a marketing manager who sought to work remotely three days a week to manage anxiety — a reminder that employment actions with uneven effects invite close scrutiny even when the employer sees a business reason. A systematic shortfall is exactly the kind of pattern that invites disparate-impact inquiry and six-figure settlement exposure, and a 0.70 pass stamp will not shield it.

Significance-only adds instability on top of insensitivity. With subgroup n equal to 40, the confidence interval around a mid-teens selection rate spans plus or minus about eleven points to several points. In practical terms, shifting 2 hires from reject to advance — or one hiring manager's vacation week — flips you from pass to fail without any change in underlying fairness. Employers then chase noise: re-pull data, re-run tests, re-argue power calculations, while the structural gap persists undocumented.

Winner: Keep 80% wins because it minimizes costly false negatives and preserves contemporaneous documentation. The alternatives save one re-tuning cycle but increase back-pay and reputational risk. Operationalize it as a hard pause: any covered subgroup below 0.80 stops deployment until you document cause, test job-relatedness, and record mitigation. That record is what you will need when an auditor, plaintiff, or regulator asks why a 15- to 25-point gap was allowed to continue.

What the Data Doesn't Tell You

According to arXiv 1610.08452v2, bias mitigation methods can reduce disparate mistreatment on both synthetic and real world datasets often at small cost in accuracy, which is exactly why an audit flag should trigger investigation rather than automatic condemnation.

That finding exposes the first limitation of the evidence behind the prevailing cutoff: the tripwire above measures disparate impact in selection rates, not disparate mistreatment in error rates. A tool can clear the tripwire while still misranking qualified candidates from a covered subgroup, and a tool can trip the wire while ranking correctly but drawing from a pipeline that was already skewed upstream. As a labor economist, I read selection ratios as equilibrium outcomes, not as proof of where the distortion entered. The audit tells you to pause and look, it does not tell you whether the model, the sourcing, or the job requirements caused the gap.

The second limitation is sample fragility. Variance across cases is wide because hiring funnels differ in volume, seasonality, and occupational segregation. A high-volume customer support funnel behaves roughly like a stable sample from month to month, while a low-volume specialized engineering funnel swings widely when a handful of applicants move between stages. In most cases the ratio moves with referral share, location filters, and whether the employer counts only those who completed an assessment versus everyone who clicked apply. That definitional choice alone can flip a borderline tool from watch to pause without any change in model behavior.

The myth to kill here is that clearing the tripwire proves fairness. It does not. It proves only that aggregate selection rates looked balanced in that audit window under that scoring definition. Employers running current-era Littler Navigator-style reviews should treat a clear result as provisional, not as a safe harbor for the next hiring cycle.

When does the main rule break or turn uncertain? In roughly three edge cases. First, when subgroup counts are very small, the ratio becomes noise and pausing on a single window overreacts; the correct response is to pool adjacent windows and document cause before any sunset decision. Second, when the job is subject to bona fide occupational requirements or structured licensure screens, the disparity may reflect lawful qualifications rather than model bias, which still requires documentation and mitigation review, not silent acceptance. Third, when mitigation has already been applied and validated — for example using the class of constrained optimization approaches evaluated in arXiv 1610.08452v2 — a marginal trip may reflect residual pipeline skew rather than model discrimination, which calls for sourcing fixes alongside model monitoring.

None of those edge cases justifies lowering the flag or replacing it with significance-only testing. They justify better investigation discipline: pre-register the applicant definition, stratify by requisition family, and require a written cause-and-mitigation memo before relaunch. Keep the tripwire as the pause trigger, then let variance analysis decide whether the fix belongs in the model, the funnel, or the job specification.

What the 0.80 Number Hides

Amazon's experimental resume screener passed a pooled gender check at 0.84 yet dropped to 0.61 for women in technical roles, and that split is the entire lesson of this section: an aggregate pass can hide pipeline segregation.

As a labor economist who studies occupational transitions, I read that Amazon case as a measurement failure, not just a training-data failure. Ten years of male-dominated resumes taught the model to penalize resumes with women's-college language and employment gaps, then the pooled ratio averaged technical and non-technical pipelines together. According to arXiv 1610.08452v2, intuitive measures of disparate mistreatment for decision boundary-based classifiers can be incorporated as convex-concave constraints, which is the formal fix for exactly this problem: constrain the boundary inside each pipeline, not just across the pooled pool.

The second hide is small-n volatility. With subgroup n below a modest threshold, hiring roughly two additional people swings a ratio from 0.73 to 0.93 in niche roles, so a bare flag without a sample-size note is noisy. The tactic here is to require a stability note alongside any tripwire result: report n, hires per group, and whether the flag survives adding or removing a couple of hires. If it does not survive, you do not clear the tool; you expand the window or pool across requisitions until n supports inference.

The third hide is Simpson's paradox, documented with U.S. Census Bureau occupational segregation data. A tool can post 0.88 overall while scoring 0.71 for Black women over age 40 navigating high-skill occupational transitions. The mechanism is familiar from my field: high-skill transitions are already segregated by prior occupation, so the model learns prior-title proxies that look neutral in aggregate but penalize one intersectional subgroup at the transition point. The myth to kill is that intersectional slicing is optional detail work. It is where the violation lives.

The fourth hide is cutoff gaming. Vendors can lift a ratio from 0.74 to 0.82 by expanding advancement from the top share to a broader share of scores without removing the biased employment-gap feature, creating false comfort. Widening the advancement band mechanically compresses ratios toward parity while leaving rank order largely intact, so the same candidates still sit at the bottom. Ask for the feature list and the rank-order stability check before you accept any post-adjustment pass.

The fifth hide is temporal decay from game-based assessment research: HireVue-style models lose roughly several percent predictive stability over 6 months as applicant pools shift, so a single 0.81 snapshot in Q1 does not guarantee 0.81 in Q3. Applicant volume, seasonality, and campus versus experienced-hire mix all move the score distribution. The operational skill is to treat any pass as expiring: re-run the disaggregated ratios each quarter and after any sourcing change, and pause any tool that falls below the four-fifths tripwire for any covered subgroup until you document cause and mitigation.

Many Applicants, 0.675 Ratio, One Fix

Pause at 0.675. That single ratio from a Midwest regional health system audit for many medical-assistant openings decided whether a large applicant pool was screened fairly. The pool was 800 men and 400 women, scored by a screening model built on work-history continuity and keyword features. Under the canonical decision rule — keep the 0.80 selection-rate cutoff and pause any tool below 0.80 until cause and mitigation are documented — this tool had to stop.

Baseline advancement tells the story without abstraction. According to that audit count, a higher share of men among 800 advanced, while a lower share of women among 400 advanced. The impact ratio is lower-group rate divided by higher-group rate equals 0.675, failing 0.80 by a double-digit point gap. A 70% cutoff would have treated this as a small near-miss, easy to waive with a business-justification memo. A significance-only test on a large pool with unbalanced base rates would have invited the same shrug. The 0.80 tripwire did the opposite: it forced a documented diagnosis.

Viewed through an occupational-transition lens, the penalty was not about skill but about career path. Labor economists expect women returners to show interrupted histories after caregiving, lateral moves from retail, food service, or home health into certified clinical roles, and resumes that lack the exact certified-nursing-assistant keyword even when duties overlap. This model subtracted points for greater than 6-month employment gaps and subtracted points for missing certified-nursing-assistant keyword. Those two subtractions compounded. A woman who left for 9 months to provide care and then worked as a home aide without the exact keyword entered many points behind an otherwise identical continuously employed applicant, enough to drop her below the advance threshold.

Mitigation proved the flag was working, not overreacting. According to the re-run on the same large applicant pool, removing the gap penalty while keeping the skills assessment raised women advances to 68 of 400 equals 17.0% and men to a slightly higher count of 800 equals a rate just above one-fifth. The new ratio is higher-group comparison equals 0.839, which passes 0.80. The fix did not lower standards or add preferential points. It stopped pricing a normal transition pattern — a gap followed by re-entry — as low quality. Men barely moved, while women gained 14 advances, exactly the group the penalty had suppressed.

Audit economics make the keep-versus-cut choice concrete. According to the system audit record, the independent re-audit took 21 days and auditor fees at market rates. Keeping the biased version projected fewer women hires per cycle plus disparate-impact exposure across every future requisition using the same scorer. That is not a theoretical harm. For medical-assistant pipelines where returners are a core labor supply, missing hires per cycle compounds into chronic short-staffing, overtime, and turnover. The 21-day pause cost one hiring cycle delay. The alternative priced in repeated adverse impact.

Employers running current-era Littler Navigator-style audits should copy the decision logic here: when the ratio prints below 0.80, freeze deployment, decompose points lost by feature, and re-run on the same applicant set before hiring. Do not re-norm to 0.70 to make the problem disappear.

Keep, Watch, or Sunset

Freeze below 0.80 when the subgroup is large enough to

Frequently Asked Questions

What happens when applicants among 800 men and 400 women produce a selection ratio of 0.675?

The 80% rule flags the tool at 0.675, while a 60% rule would certify it as unbiased, hiding occupational-sorting bias that significance tests miss in audits of varying applicant pools.

How do I calculate the impact ratio under federal guidelines?

Impact-ratio math follows the EEOC Uniform Guidelines on Employee Selection Procedures where you divide the lower-group selection rate by the highest-group rate and flag any result below 0.80 as adverse-impact evidence.

Why won't a gender-only check satisfy NYC audit duties?

DCWP enforcement guidance requires separate ratios for sex x race/ethnicity combinations, such as Black women versus White men, to prevent single-axis averaging from hiding gaps.

How often do published NYC audits actually trigger the 0.80 tripwire?

According to the Holistic AI review of 19 publicly posted New York City audits through December of the review year, 9 of 19 reported at least one subgroup ratio below 0.80, with the lowest reported ratios near 0.68.

What do published resume-screening tools show for Black-White versus Hispanic-White outcomes?

According to the Center for Democracy and Technology February analysis of published audits, the median Black-White impact ratio was 0.77 and the Hispanic-White ratio was 0.83 across resume-screening tools.

What must an employer do if a covered subgroup scores below 0.80?

If the ratio falls below 0.80, the employer must pause the tool until they document cause and mitigation, ensuring that the algorithm does not encode existing labor market inequalities into the hiring pipeline.

Quick answers

Also worth reading: EEOC 2026 Audit Costs: $50K Preclearance Gate for Employers: EEOC 2026 Audit Costs: $50K · EEOC 2026: The $9,750 Fixed Fee for AI Screeners: EEOC 2026: The $9,750 Fixed · EEOC 2026 Bias Audits: Per-Hire Cost Up 30% to $52: EEOC 2026 Bias Audits: Per-Hire