# Resume Screening Bias: 0.80 Four-Fifths Rule vs Human Override

Sarah Johnson · November 13, 2026

> AI hiring tools can pass the 0.80 four-fifths rule in aggregate yet show racially disparate impact at the job level, exposing limits of human override.

## The Evidence: Aggregate Pass, Job-Level Fail

Aggregate hiring metrics can create a false sense of compliance when AI screening tools are evaluated only at the company-wide level. Reporting summarized by techxplore.com found that AI-based game assessments used in hiring complied with the four-fifths rule in aggregate while showing racially disparate impact at the specific job level, with algorithms recommending Black and Asian applicants at disparate rates for particular positions. This convergence fact matters because an employer can appear to pass a broad screen while individual roles still produce unequal outcomes that warrant closer scrutiny.

The gap emerges because aggregate results blend applicants, qualifications, and selection patterns across many requisitions. A high-volume role with balanced outcomes can offset a smaller role where one group is consistently filtered out, masking the signal that would be visible if results were examined separately. For teams applying the reader rule, an organization-wide pass is not a stopping point: each screening stage must be evaluated on its own, and job-level disparities should be treated as evidence that the funnel requires intervention rather than reassurance.

Algorithmic monoculture compounds the risk. As described by inyourroots.com, many employers rely on the same vendor or model, allowing a biased screening pattern to replicate across firms even when each individual employer’s aggregate numbers look clean. When the same assessment logic is widely adopted, a flaw in how candidates are scored or ranked does not remain isolated to one company; it can propagate through an entire labor market segment, making job-specific disparities harder to detect through peer comparison alone.

To act on this evidence, compliance reviews should disaggregate AI screening outcomes by position and by funnel stage, not just by enterprise totals. If a job-level review reveals a disparate pattern at the AI screen, human review, or interview stage, the appropriate response is to pause reliance on the automated recommendation for that segment and require a documented human override that records what was detected, what decision was changed, and the basis for that change. This documentation creates an auditable trail showing that the organization did not treat an aggregate pass as dispositive.

The practical check is straightforward: never accept a clean aggregate report as proof that every role is fair. Pair company-wide monitoring with job-level analysis so that localized disparities in AI-driven assessments are surfaced early, investigated in context, and corrected through accountable human judgment before they become entrenched hiring patterns.

![solitary wooden footbridge forking into paths over mist covered](https://static.mm-ais.com/article-images-ai/resume-screening-bias-0-80-four-fifths-r-ai-53090054.jpg)
solitary wooden footbridge forking into paths over mist covered

## Human Override vs Auto-Reject: Options Compared

When an AI resume screen produces a disparate outcome at any stage of the hiring funnel, the architecture of human involvement determines whether an employer can defend the decision. Under the EEOC’s adverse-impact framework, described by RatedWithAI as the primary legal test for AI hiring tools in 2026, the configuration you choose either creates a checkpoint or leaves the model as the sole decision-maker. The three common approaches—auto-reject, human rubber-stamping, and documented human override—carry materially different compliance risks.

Auto-reject is the highest-risk configuration. In this model, candidates filtered out by the AI are rejected automatically, with no human checkpoint before the decision is finalized. The entire adverse-impact exposure rests on the model’s output, and the employer has no contemporaneous record showing that a person evaluated the flagged disparity or considered affected candidates individually.

Human review that merely confirms AI rankings, often called rubber-stamping, does not meaningfully reduce that exposure. If reviewers accept the system’s ordering without an independent selection decision, the process does not cure a sub-0.80 ratio. A nominal human touchpoint without independent judgment or documentation leaves the AI’s recommendation functionally intact and fails to demonstrate that the employer intervened when the screen signaled potential bias.

| Architecture | Human role | Documentation | Compliance posture |
| --- | --- | --- | --- |
| Auto-reject | None before rejection | No decision-level record tied to the flagged disparity | Highest risk: adverse-impact exposure rests entirely on the model |
| Human rubber-stamp | Confirms AI rankings without independent evaluation | Minimal or perfunctory; no evidence of a distinct decision | High risk: does not cure a sub-0.80 ratio |
| Documented human override (winner) | Re-reviews candidates affected below the ratio threshold and makes an independent decision | Records reason codes and logs each override or rejection before it is finalized | Strongest posture: creates a defensible, auditable checkpoint |

The documented human override model is the clear winner because it pairs intervention with evidence. When a stage’s results fall below the four-fifths threshold, a trained reviewer re-evaluates the affected candidates, applies job-related criteria, records standardized reason codes, and logs the final decision before any rejection is issued. This sequence establishes that a person—not just the algorithm—made the selection decision at the point of risk.

![Human Override vs Auto-Reject: Options Compared — Resume Screening Bias](https://static.mm-ais.com/article-images-pixabay/resume-screening-bias-0-80-four-fifths-r-7fae0721.jpg)

## The Numbers You Must Check Each Quarter

A quarterly monitoring cadence turns the four-fifths screen from a one-time audit into an operating control. Each quarter, calculate impact ratios separately at every funnel stage—application, AI screen, human review, and interview—rather than relying on end-of-process outcomes. A by-funnel-stage approach, as described by Classet.ai, can reveal disparities that aggregate results conceal: in the worked example later in this article, an end-to-end impact ratio of 0.50 is driven entirely by the application-to-AI-screen stage, where candidates are filtered out before a person ever reviews them, while the two downstream stages each post a neutral 1.00. To check your own funnel this way, compute the ratio at each stage from that stage's own pool and outcomes, then compare against the 0.80 threshold rather than a single end-of-process figure.

Treat 0.80 as the action threshold at every stage, not just in final hiring totals. Under 29 CFR 1607.4, a ratio below 0.80 signals potential adverse impact and should trigger immediate review of that specific stage. In a quarterly check, any sub-0.80 result should require a documented human override process for affected decisions, including a clear record of why the system’s recommendation was accepted, modified, or set aside. Stages at or above 0.80 still warrant tracking over time, because a stable ratio near the line can deteriorate as applicant volume, job requirements, or model behavior changes.

Sample size determines how much confidence you can place in each quarterly result. With fewer than roughly 30 applicants per protected group at a given stage, a single rejection can swing the ratio across the 0.80 line. Flag these low-N stages explicitly instead of treating them as proof of compliance or violation. When counts are small, widen the analysis window to combine multiple quarters or a full year, and verify your own Ns before drawing conclusions. This prevents normal variation in small pools from being mistaken for a durable pattern.

Each quarterly review should produce a consistent record: the applicant counts and impact ratios for every protected group at every stage, notes on any low-N flags, the time window used, and the disposition of any stage that fell below 0.80. Documenting these numbers on a fixed schedule creates an auditable trail that shows not only whether a disparity was detected, but also when it was detected and how the organization responded. That record is what makes the human override layer demonstrable rather than assumed.

![The Numbers You Must Check Each Quarter — Resume Screening Bias](https://static.mm-ais.com/article-images-pixabay/resume-screening-bias-0-80-four-fifths-r-8c352594.jpg)

## What the Four-Fifths Rule Does NOT Establish

The four-fifths rule, codified at 29 CFR 1607.4, is a screening threshold rather than a verdict. As nhimg.org explains, a selection rate below 80 percent of the rate for the highest-selected group "may" indicate disparity; it does not establish unlawful discrimination on its own. A sub-0.80 impact ratio functions as a trigger for closer investigation and, where appropriate, a business-necessity analysis. Treating the ratio as a pass-fail certificate risks overlooking bias that the test was never designed to resolve.

For AI resume screening, the rule’s value depends on the conditions under which it is applied. It can surface potential adverse impact quickly, but it can also produce false reassurance when sample sizes are thin, results are rolled up across roles, or multiple employers rely on similar tools. The table below summarizes when the 80 percent test breaks and when it remains useful.

| Condition | When the 80% test breaks | When it still wins |
| --- | --- | --- |
| Aggregate masking | Company-wide compliance can conceal disparate impact at the level of specific positions, as reporting summarized by techxplore.com illustrates for AI-based game assessments. | As a first-pass screen, it flags whether pooled outcomes warrant deeper, job-level review. |
| Small N | With few applicants or selections in a group, the impact ratio can swing sharply, making the 0.80 threshold an unstable signal. | It still provides a cheap, standardized check that prompts documentation and closer scrutiny of the underlying data. |
| Algorithmic monoculture | When many employers rely on the same vendor or model, as described by inyourroots.com, a passing ratio at one organization may mask systemic screening patterns across the labor market. | It remains a consistent baseline for comparing selection outcomes within a given employer’s funnel. |

Aggregate masking is especially consequential in AI hiring because a single tool may perform differently across roles, requisitions, or candidate pools. A compliant overall ratio does not establish that each job or screening context is fair. Similarly, small applicant counts can make the four-fifths calculation too volatile to interpret in isolation, while an algorithmic monoculture can normalize biased patterns that no single employer’s ratio fully reveals.

Despite these limits, the four-fifths rule remains the EEOC’s primary legal test for adverse impact in AI hiring tools, according to ratedwithai.com. Its strength is practicality: it offers a uniform, low-cost method to identify where selection outcomes diverge. Used correctly, it is not the end of compliance analysis but the starting point for determining whether a documented human review is required to examine, justify, or override a flagged screening outcome.

![What the Four-Fifths Rule Does NOT Establish — Resume Screening Bias](https://static.mm-ais.com/article-images-pixabay/resume-screening-bias-0-80-four-fifths-r-cddf78c0.jpg)

## 1,000 Applicants, Two Stages

Take 1,000 applicants: 600 in Group A and 400 in Group B. The AI screen advances 300 total—240 from A (40%) and 60 from B (15%)—so the stage impact ratio is 0.375, far below the 0.80 screen and requiring a documented human override before any rejection stands. That leaves 360 rejected from A and 340 from B, so the failure is not a small-sample artifact.

The override layer re-reviews 200 of the 700 rejected applicants and reinstates 90: 30 from A and 60 from B. Group B's cumulative advance rate rises from 15% to 30% (120 of 400), while Group A rises to 45% (270 of 600). The resulting ratio is 0.67, not 0.75; 0.75 would require Group A to remain at its original 40% rate. Because the ratio is still below 0.80, the next step is model retraining or proxy-variable ablation, not another round of overrides. The 200 re-reviewed cases are a subset of those 700, and the log should record how they were selected, because a convenience sample can hide the same disparity.

The cost of being wrong runs in both directions. Auto-rejecting applicants because a stage ratio fails, without a documented override, can discard candidates who would have survived review; overriding without a record can make a still-failing model look repaired. A false positive on compliance—treating a 0.375 ratio as acceptable because the final hire count looks balanced—leaves the failing screen in place. A false negative—rejecting every applicant after a failed ratio without review—wastes the override layer's chance to recover qualified candidates and creates a record that no human examined the disparity. The log below makes the decision auditable: it captures the ratio, the override, the reason, and the follow-up owner.

A completed override log should exist at every stage, including stages that pass, so a later audit can see both the screen and the human decision. Using the 1,000-applicant example above, an AI-screen row would read: Group A advanced 240 of 600 (40%), Group B advanced 60 of 400 (15%), impact ratio 0.375; override decision: re-review 200 of the 700 rejected and reinstate 90 (30 from A, 60 from B); reviewer and date recorded; reason and follow-up: ratio still 0.67 after override, so escalate to model retraining or proxy-variable ablation. A passing-stage row would record the same fields with the ratio at or above 0.80 and the override decision marked "no override," so the audit trail shows the screen was checked, not skipped.

## Worked Example: Run the Numbers

Illustration only: no real employer, requisition, applicant, date, or vendor is identified. Assume one 2026 high-volume customer-support requisition with 200 applicants, 100 in Group 1 and 100 in Group 2, and three downstream stages: AI screen, human review, and interview invitation.

| Stage | Group 1 | Group 2 | Impact ratio |
| --- | --- | --- | --- |
| Application to AI screen | 60/100 = 0.60 | 30/100 = 0.30 | 0.50 |
| AI screen to human review | 48/60 = 0.80 | 24/30 = 0.80 | 1.00 |
| Human review to interview | 36/48 = 0.75 | 18/24 = 0.75 | 1.00 |
| Application to interview | 36/100 = 0.36 | 18/100 = 0.18 | 0.50 |

Step 1: application to AI screen. Group 1 advances 60 of 100 (0.60); Group 2 advances 30 of 100 (0.30); impact ratio = 0.30 / 0.60 = 0.50. Step 2: AI screen to human review. Group 1 advances 48 of 60 (0.80); Group 2 advances 24 of 30 (0.80); ratio = 0.80 / 0.80 = 1.00. Step 3: human review to interview. Group 1 advances 36 of 48 (0.75); Group 2 advances 18 of 24 (0.75); ratio = 0.75 / 0.75 = 1.00. End to end, Group 1 is 36 of 100 (0.36) and Group 2 is 18 of 100 (0.18), so the overall ratio is 0.18 / 0.36 = 0.50.

The only stage below the 0.80 screen is the first one. Because the later stages are neutral, the disparity is created before human review begins; an aggregate pass at the end of the funnel would not show where the loss occurred.

The record for a flagged stage should state the ratio, the stage, the reviewer, and the reason for retaining or changing the AI output; without that, the 0.50 result is only a number.

For this illustration, the winning control is the documented human-override path: it is the path that both detects the 0.50 ratio at the application-to-AI-screen stage and leaves a record of whether the AI output was changed. The break-even trigger is the first stage whose impact ratio falls below 0.80—here, 0.50 at application to AI screen—because that is the point at which the funnel must either be documented as overridden or redesigned before the next cycle.

## Decision Rules for Your 2026 Screening Pipeline

Use these rules as the decision layer for your 2026 screening pipeline. They presuppose that the four-fifths test is run separately at application, AI screen, human review, and interview, and that a ratio below 0.80 is treated as an action trigger rather than a reporting note. The rules do not replace the underlying test; they specify what must happen after the test produces a result.

| Ratio outcome | Required action |
| --- | --- |
| Any stage below 0.80 | Documented human override before rejection: re-score, reason code, and log entry |
| Aggregate pass with shared screening vendor | Job-level ratio check for each requisition, not only a company-wide figure |
| Below 0.80 after documented overrides | Treat as a model problem: commission retraining or proxy ablation; do not scale overrides indefinitely |
| All stages at or above 0.80 with no monoculture trigger | Continue periodic monitoring and retain the check in the audit record |

If any stage's impact ratio is below 0.80, the override must be completed before any rejection at that stage. The re-score, reason code, and log create a reviewable trail showing that a person examined the flagged outcome rather than letting the stage act as an unreviewed filter. The log should identify the stage, the affected group comparison, the original score, the revised score, and the reviewer.

If the ratio passes in aggregate but you share a screening vendor with many employers, run the job-level ratio check per requisition. Monoculture risk means a vendor-level pattern can obscure a requisition-level disparity, so the check should follow the job rather than only the portfolio. Keep the vendor name, requisition identifier, and check date in the same record.

If the ratio stays below 0.80 after human overrides, escalate to retraining or proxy ablation instead of adding more overrides. Overrides are a control for individual cases; repeated reliance on them indicates that the model or its inputs need intervention. Set a review point after the retraining or ablation work so the next cycle can confirm whether the ratio moved.

Close each cycle by recording the stage, ratio, action taken, and owner. That record converts a screening metric into an auditable compliance decision and keeps the pipeline's 2026 review consistent across requisitions. Retain the record with the requisition file so later reviewers can see both the trigger and the response.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | At the application stage, calculate each group's selection rate and divide the lowest by the highest to get the impact ratio, then compare it to the 80% four-fifths threshold. | It establishes the baseline disparity measure before later screening stages can mask it. |
| 2 | Repeat the same four-fifths calculation separately at the AI screen, human review, and interview stages. | Adverse impact can appear or disappear at different points in the funnel. |
| 3 | If any stage's impact ratio falls below 80%, trigger a documented human override before the candidate is rejected. | A sub-80% ratio requires intervention rather than automatic rejection. |
| 4 | Re-score the candidate, reason-code the override decision, and log the record. | The override must be auditable, not an informal exception. |
| 5 | If no stage falls below 80%, still document every stage's impact ratio and the no-override classification. | The classification and math are what a reviewer will test. |
| 6 | Retain the stage-by-stage ratios and override records as evidence under 29 CFR 1607.4. | Compliance depends on detection and documentation of sub-80% outcomes. |

## Frequently Asked Questions

**Does passing the four-fifths rule at the company-wide level guarantee no job-level disparate impact?**

AI-based game assessments complied with the four-fifths rule in aggregate while showing racially disparate impact at the specific job level.

**Which applicant groups were recommended at disparate rates for particular positions?**

Algorithms recommended Black and Asian applicants at disparate rates for particular positions.

**Why can aggregate hiring metrics mask job-level disparities?**

Aggregate results blend applicants, qualifications, and selection patterns across many requisitions.

**How can a high-volume role affect detection of a smaller role's disparity?**

A high-volume role with balanced outcomes can offset a smaller role where one group is consistently filtered out.

**What should teams do when an organization-wide screen passes?**

Each screening stage must be evaluated on its own, and job-level disparities should be treated as evidence that the funnel requires intervention rather than reassurance.

**What compounds the risk of biased screening patterns replicating?**

Many employers rely on the same vendor or model.

## Quick answers

| Why can aggregate hiring metrics create a false sense of compliance? | Aggregate hiring metrics can create a false sense of compliance when AI screening tools are evaluated only at the company-wide level. |
| --- | --- |
| What did reporting on AI-based game assessments reveal about compliance with the four-fifths rule? | Reporting summarized by techxplore.com found that AI-based game assessments complied with the four-fifths rule in aggregate while showing racially disparate impact at the specific job level, with algorithms recommending Black and Asian applicants at disparate rates for particular positions. |
| How can aggregate results mask disparities at the job level? | Aggregate results blend applicants, qualifications, and selection patterns across many requisitions, so a high-volume role with balanced outcomes can offset a smaller role where one group is consistently filtered out. |
| How should an organization-wide pass on the four-fifths rule be treated? | An organization-wide pass is not a stopping point; each screening stage must be evaluated on its own, and job-level disparities should be treated as evidence that the funnel requires intervention rather than reassurance. |
| What risk does algorithmic monoculture pose in hiring? | Algorithmic monoculture compounds the risk because many employers rely on the same vendor or model, allowing a biased screening pattern to replicate. |

Also worth reading: **Algorithmic hiring bias ratio: 4/5ths rule flags EEOC action at 0.82 vs 0.95**: [Algorithmic hiring bias ratio: 4/5ths](https://ailaborbrain.com/blog/algorithmic-hiring-bias-ratio-45ths-rule-flags-eeoc-action-at-082-vs-095.php) · **EEOC 2026 Audit Costs: $50K Preclearance Gate for Employers**: [EEOC 2026 Audit Costs: $50K](https://ailaborbrain.com/blog/eeoc-2026-audit-costs-50k-preclearance-gate-for-employers.php) · **EEOC 2026: The $9,750 Fixed Fee for AI Screeners**: [EEOC 2026: The $9,750 Fixed](https://ailaborbrain.com/blog/eeoc-2026-the-9750-fixed-fee-for-ai-screeners.php)

### Related reading

- [California AI resume screening 2026: 80% impact test to defend audits](https://ailaborbrain.com/blog/california-ai-resume-screening-2026-80-impact-test-to-defend-audits.php)
- [Algorithmic hiring bias ratio: 4/5ths rule flags EEOC action at 0.82 vs 0.95](https://ailaborbrain.com/blog/algorithmic-hiring-bias-ratio-45ths-rule-flags-eeoc-action-at-082-vs-095.php)
- [The 80% Rule: Why Vendor Bias Badges Won't Shield Employers](https://ailaborbrain.com/blog/the-80-rule-why-vendor-bias-badges-wont-shield-employers.php)
- [2026 Screening Fails: Vector Collapse Masks Non-Linear Careers](https://ailaborbrain.com/blog/2026-screening-fails-vector-collapse-masks-non-linear-careers.php)
- [AI Screening's 12.4% Wage Compression: Mid-Career SF Tech Reality](https://ailaborbrain.com/blog/ai-screenings-124-wage-compression-mid-career-sf-tech-reality.php)
- [AI job exposure thresholds: 10% vs 50% GPT-4 task-share flags](https://ailaborbrain.com/blog/ai-job-exposure-thresholds-10-vs-50-gpt-4-task-share-flags.php)

### Latest

- [AI job exposure thresholds: 10% vs 50% GPT-4 task-share flags](https://ailaborbrain.com/blog/ai-job-exposure-thresholds-10-vs-50-gpt-4-task-share-flags.php)
- [California AI resume screening 2026: 80% impact test to defend audits](https://ailaborbrain.com/blog/california-ai-resume-screening-2026-80-impact-test-to-defend-audits.php)
- [AI Job Risk 2026: Michael Meeks’ Task-Exposure Index — 35% vs Reskill Now](https://ailaborbrain.com/blog/ai-job-risk-2026-michael-meeks-task-exposure-index-35-vs-reskill-now.php)

Canonical: https://ailaborbrain.com/blog/resume-screening-bias-080-four-fifths-rule-vs-human-override.php
Markdown: https://ailaborbrain.com/blog/resume-screening-bias-080-four-fifths-rule-vs-human-override.php/index.md