# Algorithmic hiring bias ratio: 4/5ths rule flags EEOC action at 0.82 vs 0.95

Sarah Johnson · October 14, 2026

> The 4/5ths rule flags adverse impact below 0.80. See why a 0.82 hiring ratio clears EEOC screening while 0.95 sits well clear, and how to compute it.

| Takeaway | Detail |
| --- | --- |
| The 4/5ths rule flags adverse impact when a group's selection ratio falls below 0.80 | Compute the ratio as (pass rate of the lower-scoring group) ÷ (pass rate of the highest-scoring group); a result below 0.80 is treated as an EEOC Uniform Guidelines adverse-impact flag. |
| A 0.82 selection ratio survives a first-pass EEOC inquiry | 0.82 exceeds the 0.80 threshold of the 4/5ths rule, so it clears the screening stage rather than triggering a flag. |
| A 0.95 selection ratio sits well clear of the flag zone | 0.95 is 0.15 above the 0.80 threshold, leaving substantial margin before any adverse-impact flag would apply. |
| The 0.80 threshold is a screening flag, not an enforcement verdict | The 0.82 vs 0.95 gap is the difference between surviving a first-pass EEOC inquiry and inviting a formal cause-finding; the gap is only meaningful when measured against the 4/5ths rule's threshold. |

This guide explains how the 4/5ths rule's 0.80 threshold functions as a screening flag under the EEOC Uniform Guidelines, using the 0.82 vs 0.95 selection-ratio comparison.

Readers learn to compute the ratio as the lower-scoring group's pass rate divided by the highest-scoring group's pass rate, and to treat any result below 0.80 as an adverse-impact flag.

![Algorithmic hiring bias ratio](https://static.mm-ais.com/article-images-ai/algorithmic-hiring-bias-ratio-4-5ths-rul-ai-facbb616.jpg)

## How the 4/5ths ratio actually computes

The Uniform Guidelines on Employee Selection Procedures (29 CFR Part 1607) define adverse impact with a specific fraction: if the selection rate for any race, sex, or ethnic group is less than four-fifths (80%) of the rate for the group with the highest rate, the difference is generally treated as evidence of adverse impact. That definition fixes both parts of the calculation. The numerator is the pass rate of the lower-scoring group — the share of that group's applicants who cleared the selection bar. The denominator is the pass rate of the highest-scoring group, the group that cleared at the highest rate. Divide the first by the second, and the result is the selection ratio you carry into any EEOC review.

Work an example with clean numbers. Suppose 60 of 100 applicants from Group A pass, a rate of 60%, and 45 of 100 applicants from Group B pass, a rate of 45%. The ratio is 45 ÷ 60 = 0.75. Because 0.75 falls below 0.80, the selection procedure trips the four-fifths flag for Group B. Now suppose Group B's pass rate were 50 of 100 instead: 50 ÷ 60 = 0.833, which clears the trigger. The entire inquiry turns on that single division, so compute it from raw counts — passes divided by applicants within each group — before you divide one rate by the other. Mixing applicant pools or using overall hire counts instead of group-specific pass rates is the most common way this calculation gets done wrong.

Treat the 0.80 line as a screening trigger, not a verdict. A ratio of 0.82 means the lower group's pass rate is 82% of the top group's; a ratio of 0.95 means it is 95%. Both sit above the trigger, so neither generates the flag on its own — but only 0.95 is close enough to parity that a reviewer is unlikely to open further questions. A ratio just above the line, like 0.82, survives the first-pass screen while remaining close enough to it that any change in applicant flow, cutoff score, or pass counts could push it under on the next cycle.

That fragility is the practical point. The four-fifths rule is a fixed threshold applied to a ratio computed from a single snapshot of hiring data, and fixed thresholds applied to shifting data can mislead — a pattern documented in adjacent regulatory contexts, where the arXiv analysis of algorithmic AML enforcement (arxiv.org/html/2603.04328v1) found that static classification metrics substantially overstate real-world regulatory effectiveness because thresholds calibrated once go stale as conditions drift. The check that protects you: recompute the ratio each time the applicant pool or selection procedure changes, and log the raw pass counts alongside the ratio so the trigger status is auditable rather than assumed.

One final discipline: the Guidelines' four-fifths statement is a rule of thumb for identifying adverse impact, not a legal safe harbor in either direction. A ratio above 0.80 does not immunize a procedure, and a ratio below it does not by itself end the analysis. Use the 0.80 trigger to decide when to escalate to a deeper review of the selection procedure — validation evidence, business necessity, alternative procedures — and document that escalation decision with the computed ratio attached.

![How the 4/5ths ratio actually computes — Algorithmic hiring bias ratio](https://static.mm-ais.com/article-images-ai/algorithmic-hiring-bias-ratio-4-5ths-rul-ai-f729154e.jpg)

## What the evidence does and does not show

The 4/5ths rule is a screening device, and the EEOC Uniform Guidelines treat it that way: a ratio below 0.80 triggers inquiry, not liability. What the evidence does not show is that a ratio above 0.80 clears you. The gap between a ratio that survives a first-pass inquiry and one that invites a formal cause-finding is real, but it is not a clean pass/fail line — it is a measure of how much drift your selection process can absorb before the threshold itself becomes unreliable.

The most useful recent evidence on this point comes from outside employment law. An arXiv preprint on cryptocurrency anti-money laundering enforcement (arXiv:2603.04328v1, 2026) reports that strong static classification metrics substantially overstate real-world regulatory effectiveness, because cost-sensitive enforcement thresholds destabilize as the data distribution shifts across time. That distinction matters for hiring models, because a selection ratio is a decision rule applied to a score distribution that drifts across hiring cycles, applicant pools, and job families. Check the preprint directly for the specific metrics and loss figures it reports rather than relying on any single summary.

The practical check is to stop treating a single computed ratio as a fixed property of your process. Recompute the selection ratio separately for each hiring cycle and each major job family, then compare the ratio across cycles. If the ratio moves materially from one cycle to the next without a corresponding change in the job or the applicant pool, your threshold is not stable — and a ratio that happens to sit above 0.80 in one cycle may fall below it in the next for reasons that have nothing to do with discrimination. The arXiv finding predicts exactly this pattern: static metrics look strong while the deployed decision rule drifts.

Regulators are already moving against algorithmic decision systems whose internal mechanics they cannot inspect. The DOJ has instituted enforcement actions against algorithmic pricing software providers for allegedly facilitating collusion between competitors (Reuters, via grounding). That is a different legal theory, but the enforcement posture is the same: the algorithm's output is not a defense, and a favorable aggregate metric does not insulate the operator. For hiring, the analogous move is to document the threshold-selection process itself — why 0.80, how the score distribution was validated, and what monitoring exists for drift — rather than presenting a single ratio as the whole story. Check the cited Reuters coverage for the specific providers and allegations before repeating any detail.

What the evidence does not show is that the 4/5ths rule is useless. It shows that the rule is a screening flag whose reliability depends on the stability of the underlying score distribution. A ratio that survives a first-pass inquiry and a ratio that invites a formal cause-finding can differ by a wide margin, and that margin is only meaningful if you know whether your threshold is calibrated to the current cycle or to a cycle that no longer exists.

![What the evidence does and does not show — Algorithmic hiring bias ratio](https://static.mm-ais.com/article-images-pixabay/algorithmic-hiring-bias-ratio-4-5ths-rul-f9ffe3eb.jpg)

## 82 vs 0.95: options compared

Put the two ratios side by side and the choice stops being abstract. A selection ratio of 0.82 clears the 0.80 trigger in the Uniform Guidelines on Employee Selection Procedures (29 CFR Part 1607), and a ratio of 0.95 clears it by a wider margin. The question this section answers is which treatment of each ratio you should adopt, and the comparison table below is the whole argument in one view.

| Option | Treatment of 0.82 | Treatment of 0.95 | Cost | Defensibility | Verdict |
| --- | --- | --- | --- | --- | --- |
| A | Compliant — clears 0.80 | Compliant — clears 0.80 | Lowest | Weakest | Loses |
| B | Flag — documented business-necessity review | Monitor — quarterly re-computation only | Moderate | Strongest | Wins |
| C | Clean — clears 0.80 | Clean — clears 0.80 | Lowest | Weakest | Loses on every dimension |

Option A treats 0.82 as compliant because it clears 0.80. It wins on cost — no review, no documentation, no re-run — and loses on defensibility. A plaintiff's expert will compute the distance from 0.82 to parity (1.00), show the 18-point gap, and argue the model is not job-related. The 0.80 threshold is a screening flag, not an enforcement verdict, so clearing it does not end the inquiry; it only postpones it.

Option B treats 0.82 as a flag and 0.95 as a monitor. The 0.82 triggers a documented business-necessity review: the selection procedure is examined against the job, the review is written down, and the file shows the employer did not stop at the threshold. The 0.95 triggers quarterly re-computation only, because a ratio that close to parity needs watching, not defending. This option wins on defensibility, and the added cost is the review itself — bounded, one-time, and far cheaper than a cause-finding.

Option C treats both as clean because both clear 0.80. It loses on every dimension. The 0.95 is genuinely clean and needs nothing beyond routine monitoring; the 0.82 is not, and treating it as clean forfeits the documentation that would have answered the first-pass inquiry. The error is symmetric in appearance and asymmetric in consequence: one ratio was never at risk, the other was.

The comparison also shows why the gap matters only when the lower ratio sits near the threshold. A ratio at 0.95 is far enough from 0.80 that quarterly re-computation is sufficient; a ratio at 0.82 is close enough that a small change in either pass rate can push it below the trigger. The rule that follows: compute the ratio, compare it to 0.80, and then treat anything between 0.80 and roughly 0.90 as a flag requiring a documented review, reserving monitor-only treatment for ratios at or above 0.95.

![82 vs 0.95: options compared — Algorithmic hiring bias ratio](https://static.mm-ais.com/article-images-pixabay/algorithmic-hiring-bias-ratio-4-5ths-rul-4125f50f.jpg)

## Costs and numbers that matter

The cost of misreading a ratio is not symmetric, and that asymmetry is the whole reason the 0.80 line deserves respect without being mistaken for a verdict. Treat a 0.82 ratio as clean, and you have a screening flag that survived a first pass — nothing more. If a charge follows, the employer faces early defense costs on counsel, statistical experts, and document production before any merits ruling, and the ratio itself becomes the exhibit the other side builds around. The number that looked like a pass becomes the first line of the complaint. Check your own counsel's cost estimates for a first-pass defense rather than assuming any specific figure.

Treat a 0.95 ratio as a violation, and the loss moves to a different ledger. The employer over-corrects, dilutes the model, and gives up the predictive lift that justified building the tool in the first place. That cost is measured in hiring quality — candidates who would have performed, filtered out to satisfy a threshold that was never breached — not in legal fees. No regulator fines you for a false flag; the damage is internal and slow.

The asymmetry runs the other way from what most compliance teams assume. A false clean is a contingent liability: it costs nothing until someone files, then costs a great deal at once. A false flag is a certain cost: it is paid every hiring cycle, in the quality of the people you did not see. One is a tail risk; the other is a chronic tax.

The practical check is to price both errors before you act on either. Ask what the ratio would cost if you are wrong in each direction, and whether the exposure is contingent or recurring. A ratio near the line is not a verdict; it is a prompt to run that arithmetic. The 0.82-versus-0.95 gap matters because the two numbers sit on opposite sides of a screening threshold with different downstream consequences — not because either one, standing alone, settles anything.

What the evidence does not supply is a dollar figure for either branch. The arxiv AML study (arXiv:2603.04328v1) shows that static classification metrics overstate real-world regulatory effectiveness and that fixed enforcement thresholds are fragile under temporal drift — a warning about treating any single ratio as stable, not a price list. So run the cost comparison as a method: name the exposure, name whether it is contingent or recurring, and decide on that basis. The ratio tells you where to look. The cost framing tells you what being wrong there is worth.

![Costs and numbers that matter — Algorithmic hiring bias ratio](https://static.mm-ais.com/article-images-pixabay/algorithmic-hiring-bias-ratio-4-5ths-rul-9ccda8ec.jpg)

## What the evidence does NOT establish

The 4/5ths rule is a screening device, not a legal standard — and the Uniform Guidelines themselves say so. A selection ratio below 0.80 does not automatically establish unlawful discrimination, and a ratio above 0.80 does not automatically establish compliance. The EEOC treats the threshold as a trigger for inquiry: agencies may investigate what caused the disparity, whether the procedure is job-related and consistent with business necessity, and whether less discriminatory alternatives exist. Treat the number as a flag that opens a file, not a verdict that closes one.

The rule also ignores sample size, which is where a first-pass screen can mislead you most. A ratio of 0.82 computed on a small lower-scoring group is statistically fragile — a handful of individual hiring decisions flipping either way can push the ratio across the 0.80 line in either direction, so the result may be noise. The same 0.82 computed on a very large applicant pool is a signal, because at that scale the ratio reflects a stable pattern rather than the luck of a small pool. The check: before reacting to any ratio, ask how many applicants produced it, and consider whether significance testing (such as a two-standard-deviation analysis) supports or undercuts what the raw fraction appears to show.

The rule does not address intersectional claims. It is defined for race, sex, and ethnic groups considered separately, so a workforce can post clean ratios for each protected category in isolation while a combined group — for example, women of a specific race or ethnicity — faces a compounded disparity that no single-axis ratio reveals. The practical check is to compute selection rates not only for each protected category but for the intersections relevant to your applicant pool, and to document what those combined rates show even though the Guidelines' threshold language does not explicitly require it.

Finally, the rule says nothing about calibration over time. Research on algorithmic enforcement systems — including the arXiv study of cryptocurrency anti-money-laundering models — reports that strong static metrics substantially overstate real-world regulatory effectiveness because decision thresholds become miscalibrated as conditions drift. The analogous risk in hiring is treating a one-time ratio as a permanent clearance: applicant pools, recruiting channels, and screening procedures change, and a ratio that passed review last cycle can fail this one. The check is periodic recomputation on current data, not a single archived calculation.

What the evidence does not establish, then, is that the 0.80 line is a compliance floor or ceiling. It is a first-pass screen whose output — like the loss-based evaluation frameworks the AML literature argues for — should feed a broader inquiry into sample size, statistical significance, intersectional effects, and temporal stability before anyone treats the number as dispositive.

![What the evidence does NOT establish — Algorithmic hiring bias ratio](https://static.mm-ais.com/article-images-pixabay/algorithmic-hiring-bias-ratio-4-5ths-rul-7a82557c.jpg)

## 1,200 applicants, two ratios

Start with the funnel. 1,200 applicants enter the top of the process; screening narrows the pool to 400 who reach the algorithmic scoring stage. The model passes 240 of those 400, an overall pass rate of 60%. That aggregate number is where most audits stop — and it is exactly where the adverse-impact question begins, because the overall rate tells you nothing about how the model treats each group inside the pool.

Split the 400 at the scoring stage. Group A, the highest-scoring group, contributes 200 applicants, of whom 140 pass the model — a pass rate of 70.0%. Group B, the lower-scoring group, also contributes 200 applicants, of whom 115 pass — a pass rate of 57.5%. Check the arithmetic before you trust it: 140 ÷ 200 = 0.700, and 115 ÷ 200 = 0.575. Both group pass rates are computed on the same denominator, 200, so the comparison is clean.

Now apply the selection-ratio definition: divide the lower-scoring group's pass rate by the highest-scoring group's pass rate. Here that is 0.575 ÷ 0.700, which works out to 0.821. Under the EEOC Uniform Guidelines' four-fifths convention, a ratio below 0.80 is the trigger for inquiry. At 0.821, this funnel sits just above the line — a first-pass screening result that does not flag, but with almost no margin.

The lesson of the worked example is how little separates a defensible ratio from a flagged one. If Group B had passed 155 instead of 115 applicants — a difference of 40 people out of 200 — the ratio would move dramatically. Run the same division on any variant you test: recompute both pass rates, divide, and compare to 0.80. Small changes in who reaches the scoring stage move the ratio more than small changes in who passes it, so audit the funnel inputs, not just the model outputs.

Treat the computed ratio as a screening flag, not an enforcement verdict. A ratio above 0.80 survives the initial Uniform Guidelines screen; it does not certify the model as compliant. The recent arXiv work on algorithmic enforcement systems in cryptocurrency AML shows why: static classification metrics substantially overstate real-world regulatory effectiveness, and miscalibrated decision rules — not declining accuracy — drive the losses. A single snapshot ratio is a static metric. Re-run this computation on each hiring cycle, and if the ratio dips below 0.80, treat it as the trigger for a formal cause-finding inquiry rather than a verdict already rendered.

## Worked Example: Run the Numbers

Illustration: a 2026 hiring round at a mid-sized employer. The HR analytics team reviews 200 applicants: 100 in the highest-scoring group and 100 in a lower-scoring group. The highest-scoring group has 60 selected; the lower-scoring group has 49 selected. These inputs are hypothetical, used only to run the arithmetic end to end.

Step 1: compute the pass rate for the highest-scoring group. 60 ÷ 100 = 0.60. Step 2: compute the pass rate for the lower-scoring group. 49 ÷ 100 = 0.49. Step 3: apply the reader rule’s ratio calculation to these pass rates: 0.49 ÷ 0.60 = 0.816666…, which rounds to 0.82. Because 0.82 is above 0.80, the ratio clears the numeric threshold in this example. The highest-scoring group is the winner, with a pass rate 0.11 higher than the lower-scoring group.

The break-even trigger is the lower group’s pass rate that makes the ratio exactly 0.80. Multiply the highest-group pass rate by 0.80: 0.60 × 0.80 = 0.48. If the lower-scoring group’s pass rate falls to 0.48 — meaning 48 selected out of 100 instead of 49 — the ratio becomes 0.48 ÷ 0.60 = 0.80. At 0.47, the ratio is 0.47 ÷ 0.60 = 0.7833, which is below 0.80. The margin between clearing and not clearing is one selection out of 100 applicants.

Here is the calculation in table form, labeled as an illustration.

| Group | Applicants | Selected | Pass rate | Ratio step |
| --- | --- | --- | --- | --- |
| Highest-scoring | 100 | 60 | 0.60 | — |
| Lower-scoring | 100 | 49 | 0.49 | 0.49 ÷ 0.60 = 0.82 |
| Break-even trigger | 100 | 48 | 0.48 | 0.48 ÷ 0.60 = 0.80 |

The worked example shows that a single selection difference can move the ratio from 0.82 to 0.80. The winner is the highest-scoring group; the break-even trigger is a lower-group pass rate of 0.48. Run the same steps with your own applicant counts to find your break-even trigger before you rely on the ratio for any decision.

## Decision rules for the 0.80-to-0.95 band

Once the selection ratio is computed, the number only becomes useful when it maps to an action. The decision rules below treat the 0.80 trigger as the floor of a graduated response, not a binary pass/fail verdict — consistent with the Uniform Guidelines' framing of the four-fifths rule as a screening device that triggers inquiry rather than a finding of discrimination.

**Rule 1 — ratio below 0.80:** treat this as an EEOC adverse-impact flag. Before go-live, document business necessity and job-relatedness for the selection procedure, and do not deploy without a written justification on file. The written record should identify which group's pass rate drives the ratio, what validated job requirement the procedure measures, and what alternatives were considered. A ratio in this zone is the one that invites a formal cause-finding if deployed undocumented.

**Rule 2 — ratio from 0.80 up to but not including 0.90:** treat this as a monitoring-zone flag. The point estimate clears the trigger, but a single cycle proves little. Recompute the ratio every hiring cycle and document the trend, not just the current value. A ratio that drifts from 0.88 to 0.81 across three cycles is a materially different risk than one holding steady at 0.88, even though both sit in the same band at any single snapshot.

**Rule 3 — ratio from 0.90 up to but not including 0.95:** treat this as a watch item. Recompute per cycle and flag any single-cycle drop of more than 5 ratio points. A fall from 0.93 to 0.87 moves the procedure from watch item into the monitoring zone in one cycle, and the trend line — not the crossing of any single threshold — is what should prompt a documented review of the selection procedure's inputs.

| Ratio band | Classification | Required action |
| --- | --- | --- |
| Below 0.80 | Adverse-impact flag | Document business necessity and job-relatedness before go-live; no deployment without written justification |
| 0.80 to under 0.90 | Monitoring-zone flag | Recompute per hiring cycle; document the trend, not just the point estimate |
| 0.90 to under 0.95 | Watch item | Recompute per cycle; flag any single-cycle drop of more than 5 ratio points |

Two cautions govern all three bands. First, the bands are decision aids, not safe harbors: a ratio above 0.80 does not by itself clear a procedure, because the Guidelines permit inquiry into practices whose ratios sit above the trigger. Second, the arXiv work on algorithmic enforcement in digital-asset compliance shows why fixed thresholds decay — temporal non-stationarity destabilizes cost-sensitive decision rules, generating persistent excess regulatory loss even when the underlying model's predictive accuracy holds. The same mechanism applies here: a ratio computed once at launch is a snapshot of a moving system, which is why every band above ties its action to per-cycle recomputation rather than a one-time certification.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Pull the pass rate of the lower-scoring group and the pass rate of the highest-scoring group from your selection data, then divide the first by the second to produce the selection ratio. | This single quotient is the only input the 4/5ths rule needs; without it you cannot tell a screening flag from a clean result. |
| 2 | Compare the computed ratio against the 0.80 threshold: a result below it is an EEOC Uniform Guidelines adverse-impact flag, so document the business necessity before go-live. | The threshold is a screening flag, not an enforcement verdict — but it must be answered with a documented justification, not silence. |
| 3 | If the ratio lands between 0.80 and 0.95, log it as a monitoring zone and schedule quarterly re-computation rather than treating it as a clean bill of health. | A ratio in this band survives a first-pass EEOC inquiry but leaves no margin to assume the model is safe. |
| 4 | Re-check the 0.82 selection ratio against the 0.80 threshold and record that it clears the screening stage. | 0.82 exceeds the 4/5ths threshold, so it does not trigger a flag — but the margin is thin enough to warrant the monitoring cadence. |
| 5 | Re-check the 0.95 selection ratio and note the margin it holds above the 0.80 threshold. | 0.95 sits well clear of the flag zone, which is the difference between surviving a first-pass EEOC inquiry and inviting a formal compliance review. |
| 6 | Re-run the ratio computation on the same lower-scoring and highest-scoring pass rates at each quarterly checkpoint, and re-document business necessity whenever the result drops below 0.80. | The 0.82 vs 0.95 gap is not a one-time verdict; it is a standing obligation to recompute as selection data changes. |

## Frequently Asked Questions

**At what selection ratio does the 4/5ths rule flag adverse impact?**

The 4/5ths rule flags adverse impact when a group's selection ratio falls below 0.80.

**How is the 4/5ths selection ratio calculated?**

Compute the ratio as (pass rate of the lower-scoring group) ÷ (pass rate of the highest-scoring group).

**What happens to a selection ratio that exceeds the 4/5ths threshold?**

A 0.82 selection ratio survives a first-pass EEOC inquiry because it exceeds the 0.80 threshold of the 4/5ths rule, so it clears the screening stage rather than triggering a flag.

**How much margin does a 0.95 selection ratio have before an adverse-impact flag applies?**

A 0.95 selection ratio sits 0.15 above the 0.80 threshold, leaving substantial margin before any adverse-impact flag would apply.

**Is the 0.80 threshold an enforcement verdict or a screening flag?**

The 0.80 threshold is a screening flag, not an enforcement verdict.

**What does the gap between a 0.82 and a 0.95 selection ratio actually represent?**

The 0.82 vs 0.95 gap is the difference between surviving a first-pass EEOC inquiry and inviting a formal cause-finding, and the gap is only meaningful when measured against the 4/5ths rule's threshold.

## Quick answers

| What does the 4/5ths rule flag? | The 4/5ths rule flags adverse impact when a group's selection ratio falls below the 4/5ths threshold. |
| --- | --- |
| How is the 4/5ths ratio computed? | Compute the ratio as the lower-scoring group's pass rate divided by the highest-scoring group's pass rate. |
| What happens to the 0.82 selection ratio? | The 0.82 selection ratio survives a first-pass EEOC inquiry because it exceeds the 4/5ths threshold, so it clears the screening stage rather than triggering a flag. |
| What does the 0.95 selection ratio do? | The 0.95 selection ratio sits well clear of the flag zone, leaving substantial margin before any adverse-impact flag would apply. |
| What is the 0.82 vs 0.95 gap? | The 0.82 vs 0.95 gap is the difference between surviving a first-pass EEOC inquiry and inviting a formal cause-finding, and it is only meaningful when measured against the 4/5ths rule's threshold. |

Also worth reading: **EU AI Act 2026: documenting algorithmic hiring-system controls**: [EU AI Act 2026: documenting](https://ailaborbrain.com/blog/eu-ai-act-2026-documenting-algorithmic-hiring-system-controls.php) · **Hiring bias audits: Threshold in 12 days vs retrain in 9 weeks**: [Hiring bias audits: Threshold in](https://ailaborbrain.com/blog/hiring-bias-audits-threshold-in-12-days-vs-retrain-in-9-weeks.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)

### Related reading

- [EU AI Act 2026: documenting algorithmic hiring-system controls](https://ailaborbrain.com/blog/eu-ai-act-2026-documenting-algorithmic-hiring-system-controls.php)
- [Which AI Hiring Tools Need a Bias Audit: NYC Local Law 144 Rules for 2026](https://ailaborbrain.com/blog/which-ai-hiring-tools-need-a-bias-audit-nyc-local-law-144-rules-for-2026.php)
- [Hiring bias audits: Threshold in 12 days vs retrain in 9 weeks](https://ailaborbrain.com/blog/hiring-bias-audits-threshold-in-12-days-vs-retrain-in-9-weeks.php)
- [Resume screening bias: 89% accurate fails 80% rule conditional keep](https://ailaborbrain.com/blog/resume-screening-bias-89-accurate-fails-80-rule-conditional-keep.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)
- [EEOC 2026 Bias Audits: Per-Hire Cost Up 30% to $52](https://ailaborbrain.com/blog/eeoc-2026-bias-audits-per-hire-cost-up-30-to-52.php)

### Latest

- [Which AI Hiring Tools Need a Bias Audit: NYC Local Law 144 Rules for 2026](https://ailaborbrain.com/blog/which-ai-hiring-tools-need-a-bias-audit-nyc-local-law-144-rules-for-2026.php)
- [Who Gets the Overtime Shift: 2026 Fair Workweek Gap — 24% Audit or Trust](https://ailaborbrain.com/blog/who-gets-the-overtime-shift-2026-fair-workweek-gap-24-audit-or-trust.php)
- [EU AI Act 2026: documenting algorithmic hiring-system controls](https://ailaborbrain.com/blog/eu-ai-act-2026-documenting-algorithmic-hiring-system-controls.php)

Canonical: https://ailaborbrain.com/blog/algorithmic-hiring-bias-ratio-45ths-rule-flags-eeoc-action-at-082-vs-095.php
Markdown: https://ailaborbrain.com/blog/algorithmic-hiring-bias-ratio-45ths-rule-flags-eeoc-action-at-082-vs-095.php/index.md