# AI Hiring Bias Audits: NYC's Four-Fifths Rule—Document or Defend?

Sarah Johnson · October 20, 2026

> Learn how NYC’s Four-Fifths Rule applies to AI hiring, how to calculate selection rates, document adverse impact, and respond to bias audit findings.

| Takeaway | Detail |
| --- | --- |
| Test selection rates at every AI-influenced hiring stage | Calculate and retain protected-group selection rates for each AI-touched stage, including early sourcing filters, recruiter prescreens, resume parsing, and other selection steps. |
| Treat 0.80 as the adverse-impact review threshold | A selection-rate ratio below 0.80 relative to the highest group’s rate is evidence of adverse impact and should trigger investigation and documentation. |
| Do not let a passing stage offset a failing stage | If any AI-influenced stage has a ratio below 0.80, do not defend the model as lawful merely because other stages meet the threshold. |
| Separate a ratio finding from a legality verdict | A four-fifths ratio below 0.80 is a defense-triggering signal requiring review, not by itself a standalone determination that the hiring practice is unlawful. |

This guide explains how NYC employers can document four-fifths selection-rate ratios across AI-influenced hiring stages. It provides a stage-level framework for identifying ratios below 0.80, triggering an adverse-impact defense, and avoiding premature legal conclusions.

![Rain slicked Manhattan streets between glass corporate towers pale](https://static.mm-ais.com/article-images-ai/ai-hiring-bias-audits-nyc-s-four-fifths-ai-b17bac10.jpg)
Rain slicked Manhattan streets between glass corporate towers pale

## How four-fifths works at each AI stage

The four-fifths test operates at the level of a selection rate. For each protected group, divide that group’s selection rate by the highest selection rate observed among the groups: *group selection rate ÷ highest group selection rate*. A result below **0.80** is generally regarded as evidence of adverse impact, according to TrueStep HR’s 2026 explanation of the rule. For a 2026 NYC AI hiring bias audit, retain that ratio for every AI-influenced selection stage, and treat a sub-0.80 result as a signal requiring legal and control review rather than a conclusion about the model’s lawfulness.

Apply the test to every selection step touched by AI, including steps that occur before a formal application. The American Institute for Employment Readiness expressly identifies early sourcing filters, recruiter pre-screens, and resume parsing as points where the four-fifths rule may apply. This pre-application scope matters because an automated sourcing exclusion or resume-ranking decision can affect who reaches the formal hiring process even if no application is yet on file. The audit inventory should therefore start with the first AI-assisted screen and continue through each later stage where applicants are selected, filtered, advanced, or rejected.

The first method check is to list every AI-touched stage before calculating anything. For each stage, identify the people exposed to it, the specific selection event, the system or tool involved, and the resulting disposition. A useful record distinguishes among sourcing, recruiter screening, resume parsing, assessment scoring, interview-stage ranking, and final selection. Human review does not remove a stage from the inventory if AI output influenced the decision; instead, the record should identify where that influence occurred and what the downstream decision was.

At each stage, retain the applicant count and selection count for every protected group. Calculate the group’s selection rate as *selections ÷ applicants*, then calculate the stage ratio as *group selection rate ÷ highest group selection rate at that stage*. Keep the stage, denominator, and selection event attached to the result. Before accepting the arithmetic, confirm that each group has a valid applicant denominator, that the numerator matches the defined selection event, and that the comparison group is the highest-rate group for that same stage rather than for the overall funnel.

Do not collapse these stage-level results into a single aggregate rate. An overall ratio can conceal a sub-0.80 result at an early filter, while a later-stage ratio can appear acceptable because a large pool advanced to it. The defensible audit record is a stage-by-stage matrix: each AI-influenced stage, each protected group’s applicants and selections, the resulting selection rate, the four-fifths ratio, and the source data used in the calculation. Retain that matrix with model-version, threshold, and workflow-change records so that any ratio below 0.80 can be traced to the decision step that produced it.

![How four-fifths works at each AI stage — AI Hiring Bias Audits](https://static.mm-ais.com/article-images-ai/ai-hiring-bias-audits-nyc-s-four-fifths-ai-19eea1e1.jpg)

## Evidence: what 2026 primers converge

Several 2026 primers converge on the same operational point: the four-fifths rule is an adverse-impact signal that requires careful, stage-specific analysis. It is not a shortcut to declaring an AI-assisted hiring process lawful. The practical check for an audit team is therefore whether the underlying selection-rate data are complete, correctly calculated, and retained for the relevant screening stages.

| Source | Date | Converging point |
| --- | --- | --- |
| hirebetter.org | Sep 15, 2026 | Describes the rule as a measure of adverse impact and explains that its use requires specific hiring data. |
| TrueStep HR | Jun 2, 2026 | Identifies 80% as the federal four-fifths trigger for evidence of adverse impact. |
| assessall.com | Sep 16, 2026 | Explains that clearing four-fifths means the absence of one adverse-impact trigger, not a finding of lawfulness. |
| FirstPanel | Aug 23, 2026 | Warns that nearly correct implementation can produce a confident but incorrect legal conclusion. |

The first check is evidentiary. hirebetter.org frames the rule around both what it measures and the data needed to evaluate it, so a defensible audit cannot rely on a final hire-rate comparison alone. For every AI-influenced screening stage, the team should retain the protected-group selection rates used in the comparison, identify the denominator for each rate, and preserve the records needed to reproduce the calculation. A missing stage-specific data set should be treated as an audit limitation rather than silently treated as a passing result.

The second check is interpretation. TrueStep HR’s primer identifies the federal trigger, while assessall.com draws the critical distinction between clearing that trigger and establishing lawfulness. Accordingly, a ratio at or above the trigger should not become an affirmative legal defense. Equally, a result below 0.80 should trigger review and evidence preservation concerning the affected stage; it should not be presented as a standalone verdict that the employer has violated the law.

The third check is calculation quality. FirstPanel’s warning is especially relevant to AI hiring because automated workflows can make an apparently polished audit number look more reliable than its inputs. Reviewers should test the group labels, numerator and denominator definitions, exclusions, stage boundaries, and source records, then require a second reviewer to reproduce the result from the retained data. A conclusion is stronger when another person can follow the same records and arrive at the same ratio without relying on the model’s output.

As the takeaway table above states, 0.80 is best treated as a defense-triggering adverse-impact threshold for scrutiny, not as a lawful-pass certificate; the decision rules later in this guide apply that distinction stage by stage.

![Evidence: what 2026 primers converge — AI Hiring Bias Audits](https://static.mm-ais.com/article-images-pixabay/ai-hiring-bias-audits-nyc-s-four-fifths-09f764e2.jpg)

## Options compared: document vs defend

For 2026 New York City AI hiring bias audits, the key decision is not whether to document or defend; it is whether documentation precedes any defense. The defensible operating posture is to compare three approaches against the audit’s evidence needs: continuous step-level documentation, defense after a failed screening comparison, and reliance on an aggregate pass rate or no audit file. Of these, continuous documentation is the strongest approach because it creates an auditable record of what happened at each AI-influenced stage.

| Option | What it means | Evidence fit | Verdict |
| --- | --- | --- | --- |
| Document continuously | Compute per-stage ratios, retain the underlying counts, and log remediation. | Matches the adverse-impact trigger and the pre-application scope of the audit. | Winner |
| Defend after failure | Rely on validity and job-relatedness only after a ratio falls below 0.80. | A necessary contingency, not a substitute for ongoing monitoring. | Runner-up |
| Ignore or use an overall pass | Use one aggregate rate or maintain no audit file. | Conflicts with step-level testing and the rule’s treatment as a signal rather than a safe harbor. | Reject |

Continuous documentation wins because it answers operational questions an aggregate analysis cannot. The audit file should identify each protected group, the number selected and the number evaluated at each AI-touched screening stage, the resulting stage-level comparison, any review of the process, and the corrective action taken. The AIEH’s “The Four-Fifths Rule (80% Rule)” specifically places the rule at every selection step, including early sourcing filters, recruiter prescreens, and résumé parsing. That makes a passing overall rate an inadequate substitute for stage-specific records.

Defending after failure remains necessary, but it is a response to a documented signal rather than the primary control. TrueStep HR reports that a selection rate below four-fifths of the highest group’s rate is generally regarded as evidence of adverse impact. Once that signal appears, the organization should preserve the relevant data, examine the criterion or AI workflow involved, assess job-relatedness, and record the remediation decision. Validation evidence may support a response; by itself, it cannot erase the need to monitor every stage.

An aggregate pass or missing file is the weakest option because both weaken reviewability. An overall rate can conceal a disparity concentrated in one screening step, while an absent file leaves the organization unable to show what was measured or whether an identified concern was addressed. AssessAll’s analysis of Uniform Guidelines language makes the distinction clear: clearing four-fifths removes one adverse-impact trigger but does not establish lawfulness. Accordingly, the review file should support a sequence of documented checks, not a single conclusion that the process is safe because one summary figure appears acceptable.

The practical control is simple: begin with stage-level records, retain the counts used to produce each comparison, and open a documented review whenever a comparison falls below 0.80. A defense memorandum can then be attached to the underlying audit record, with its scope, evidence, limitations, and remediation explicitly stated. This structure distinguishes monitoring from advocacy and keeps the organization’s response tied to the process that produced the signal.

![Options compared: document vs defend — AI Hiring Bias Audits](https://static.mm-ais.com/article-images-pixabay/ai-hiring-bias-audits-nyc-s-four-fifths-5f164ef9.jpg)

## Costs and numbers that matter

For a 2026 NYC AI hiring bias audit, build the budget around the work required to produce an auditable record, not around a single model review. The unit of work is the stage. For every AI-touched stage, budget for: the number of stages; the number of applicants in each protected group entering that stage; the number from each group selected or advanced; and the frequency of the audit. Those are the raw inputs the analysis needs. The budget sheet should also assign an owner to extract, reconcile, and retain the data, because a calculation that cannot be reconstructed is not a reliable audit record.

| Budget quantity | Minimum record to retain | Cost driver |
| --- | --- | --- |
| Number of AI-touched stages | Stage inventory, system or vendor, owner, and decision rule | Each stage requires separate data extraction, validation, and review |
| Applicants per group per stage | Group-level stage-entry counts and applicable definition | Larger populations require more complete reconciliation and quality control |
| Selections per group per stage | Group-level advancement or selection counts tied to the stage | Counts must be traceable to source records, not reconstructed from memory |
| Audit frequency | Defined cadence, refresh dates, and change-triggered reviews | Repeating the work after data, model, vendor, or process changes |

Stage count is the main multiplier. A three-stage process does not create one review task; it creates three. Early sourcing filters, recruiter pre-screens, and resume parsing each add review work, as reflected in the AIEH primer’s treatment of AI use across selection steps. Budget for stage-level validation of inputs, exclusions, outcomes, timestamps, and control totals, plus staff time to explain who counted each applicant and why. Include integration work where hiring-platform, ATS, vendor, and HRIS records do not align; otherwise the audit may spend time reconciling systems rather than assessing the hiring process.

Plan for two wrong-both-ways errors. A false clean audit occurs when missing or inconsistent data is treated as a clean result, leaving the organization without a defensible explanation if the process is challenged. That creates a defense gap: the organization cannot show what was tested, when it was tested, or which records supported the result. The FirstPanel primer warns that easy-to-implement versions can produce confident but incorrect legal conclusions, so the budget should include independent quality checks rather than accepting a dashboard export at face value.

A false alarm occurs when thin group cells are treated as reliable evidence of a system problem. That can trigger unnecessary retraining, vendor remediation, workflow changes, or broader legal review. Reserve capacity for a defined escalation review that checks cell size, missingness, coding consistency, and whether the apparent result is stable under a reasonable re-performance of the calculation. AIEH’s May 4, 2026 primer supports examining every selection step; the cost control is to make that examination staged, documented, and proportionate rather than skipping the record or overreacting to noise.

![Costs and numbers that matter — AI Hiring Bias Audits](https://static.mm-ais.com/article-images-pixabay/ai-hiring-bias-audits-nyc-s-four-fifths-b245b1a7.jpg)

## What the evidence does NOT establish

A ratio at or above 0.80 does not establish that an AI-influenced hiring process is lawful. It shows only that the relevant comparison has not crossed the adverse-impact trigger. As the Uniform Guidelines discussion summarized by AssessAll explains, clearing four-fifths is “the absence of one trigger, not a finding of lawfulness.” TrueStep HR likewise characterizes a sub-0.80 result as evidence of adverse impact rather than a complete legal determination. The defensible operational rule is therefore narrower: a passing ratio closes one defense-triggering checkpoint, but it does not establish compliance with the New York City Human Rights Law, the employment discrimination laws enforced by the EEOC, or any other applicable standard.

That distinction matters even when every stage clears the threshold. A lawful assessment still requires examination of the screening process, the job relatedness of criteria, consistency of administration, available accommodations, and the record supporting the employer’s explanation for any disparity. A four-fifths calculation cannot answer those questions. FirstPanel cautions that implementing the rule “almost right” can produce a confident but incorrect legal conclusion; the practical check is to retain the underlying stage-level records, identify the affected group, and explain what happens next when the ratio is below 0.80. A passing calculation should not be used as the sole support for declaring the process lawful.

The available sources also does not establish NYC-specific filing deadlines, required audit formats, certification procedures, or substantive legal tests. Those items must be verified against the current statutory text, the law’s effective date, and the filing instructions issued by the responsible NYC agency or other enforcing body. Before making a compliance claim, check the official source rather than inferring a requirement from a secondary primer. Preserve the version of the materials relied upon, including its publication or update date, and record the source, retrieval date, and responsible reviewer in the audit file.

| Edge case | When four-fifths breaks | When it still wins |
| --- | --- | --- |
| Tiny group counts | The ratio is fragile because a small change in the numerator or denominator can materially change the result. Check the sample size, cohort definition, and calculation before treating the result as conclusive. | It remains usable as an adverse-impact signal when the underlying records are complete, the comparison is transparent, and the reviewer documents the uncertainty rather than presenting the ratio as proof of lawfulness. |
| Zero or unavailable comparison rate | No valid ratio can be calculated, so four-fifths neither passes nor fails. Recheck eligibility, data capture, stage definitions, and the denominator. | It does not “win”; it becomes a data-quality exception that must be resolved and retained in the audit record. |

Finally, the evidentiary standard in the thesis is deliberately asymmetric. A ratio below 0.80 is a defense-triggering signal that requires an opened inquiry; it is not, by itself, a verdict that the employer violated the law. The defensible file therefore documents each protected group’s selection rate at every AI-influenced screening stage, retains the calculation, and records the response to any triggered review. A result at or above 0.80 may support an internal checkpoint, but only verified current law and the full process record can support a broader claim of compliance.

![What the evidence does NOT establish — AI Hiring Bias Audits](https://static.mm-ais.com/article-images-pixabay/ai-hiring-bias-audits-nyc-s-four-fifths-4ae80d26.jpg)

## 0.60 ratio worksheet

For a 2026 NYC AI hiring bias audit, begin with an AI resume parser. Confirm that the stage is AI-touched and occurs before an application is submitted. The AIEH primer (aieh.com) identifies resume parsing and other early-stage screening processes as selection steps covered by the four-fifths rule, so this stage should not be omitted from the audit record.

Next, validate the denominator and numerator. The denominator must be the number of applicants in each protected group who reached the stage, and the numerator must be the number from that group who advanced beyond it. These are applicant-level selection counts, not hiring counts. A parser that advances applicants who never reach the full application stage should still be evaluated on the records it actually screened.

For the hypothetical AI resume parser, Group A has 100 applicants and 50 advance, producing a selection rate of 50.0%. Group B has 100 applicants and 30 advance, producing a selection rate of 30.0%. The four-fifths ratio is 30.0% ÷ 50.0% = 0.60. Because 0.60 is below the 0.80 trigger, this stage generates a defense-triggering adverse-impact signal. As TrueStep HR explains, a selection rate below four-fifths of the highest group’s rate is generally treated as evidence of adverse impact.

| Stage | Group | Applicants | Selected | Selection rate | Ratio to highest rate | Flag and action |
| --- | --- | --- | --- | --- | --- | --- |
| AI resume parser | Group A | 100 | 50 | 50.0% | 1.00 | Highest observed rate |
| AI resume parser | Group B | 100 | 30 | 30.0% | 0.60 | Below 0.80: open defense file |

The 0.60 result should trigger documentation, not an unsupported assertion of compliance. Passing other stages does not erase a sub-0.80 ratio at an AI-touched stage. Nor does the ratio alone establish unlawfulness. It identifies a condition that requires a documented response.

Open a defense file containing job-relatedness evidence for the parser, the validation methodology, the underlying counts, and a remediation log. Preserve the worksheet for every protected group and every AI-influenced screening stage. The defensible record shows what was measured, where the trigger occurred, what evidence was collected, and what corrective action followed.

## Decision rules: document vs defend

The 2026 audit posture should follow a single set of if/then decision rules. First, identify every stage where AI influences applicant evaluation, ranking, screening, rejection, or advancement. For each stage, retain the number of applicants in each protected group and the number selected or advanced from each group. These records are the basis for the audit file and should be preserved in a form that permits stage-level review rather than only reporting a company-wide outcome.

**If an AI-touched stage has no protected-group selection counts, document the gap before deployment and do not defend the model on design intent alone.** An explanation that the tool was intended to be fair, uses a reputable vendor, or passed an accuracy test cannot substitute for selection-rate evidence. Before the stage is used, identify the relevant applicant population, define the selection event, assign the stage’s decision to the appropriate system and owner, and begin collecting the counts needed to evaluate disparate outcomes. A stage without this evidence is an open documentation item, not a passing stage.

**If any stage’s selection-rate ratio is below 0.80, start a defense-and-remediation file for that stage.** The file should preserve the calculation inputs, identify the affected group and comparison group, describe the AI’s role in the decision, and record the proposed response, such as adjusting a threshold, changing a ranking rule, adding human review, or retesting. Do not average the deficient ratio with passing stages. A later stage cannot erase an adverse-impact signal created at an earlier screening point, and passing results elsewhere in the hiring process do not answer the specific problem at the flagged stage.

The defense file should also distinguish a signal from a final legal conclusion. TrueStep HR describes a ratio below four-fifths as evidence of adverse impact, while AssessAll emphasizes that clearing four-fifths means only that one trigger did not appear. Accordingly, the file should state the measured result plainly, preserve any qualifications concerning the data or the process, and explain what remediation or further legal analysis is being pursued. The defensible statement is that the audit identified a stage requiring review—not that the model is unlawful merely because the threshold was crossed, or lawful merely because another stage passed.

**If every AI-touched stage has a ratio at or above 0.80, retain the audit file and state only that no four-fifths trigger appeared.** The record should still include the stage inventory, protected-group counts, calculations, data-quality notes, and audit date. The conclusion should not say that the process is lawful, compliant, bias-free, or cleared by the four-fifths rule. As AssessAll explains, passing the threshold shows the absence of one trigger rather than establishing lawfulness. This wording preserves the evidentiary value of the audit while avoiding a broader assurance that the data, implementation, job relevance, or rest of the hiring process cannot support.

Apply these rules without substituting a portfolio score for stage-level review. The operating check is simple: missing counts keep a stage in documentation status; a ratio below 0.80 opens a defense-and-remediation file; and ratios at or above 0.80 permit only the limited conclusion that no four-fifths trigger appeared. This makes the audit file both reproducible and appropriately bounded.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Map every AI-influenced hiring stage used in NYC, including early sourcing filters, recruiter prescreens, resume parsing, and other selection steps. | The audit must cover the full selection process; reviewing only the final hiring stage can conceal discriminatory exclusion earlier in the funnel. |
| 2 | Calculate and retain selection rates for each protected group at every mapped stage, documenting the data source, population, and calculation for each rate. | Contemporaneous records provide the evidence needed to identify disparities and explain how AI-assisted decisions were made. |
| 3 | Compare each stage’s protected-group selection rates against the highest group’s rate using NYC’s four-fifths adverse-impact review threshold. | A failing comparison identifies the specific stage requiring scrutiny rather than allowing a stronger result elsewhere to conceal a weaker stage. |
| 4 | For any stage below the four-fifths threshold, open a validity and job-relatedness defense file before deployment. | The stage-level finding triggers a substantive defense review; passing stages do not independently establish that the AI-assisted hiring process is lawful. |
| 5 | In that file, document the business necessity, job-relatedness, validation results, and less discriminatory alternatives considered for the AI-assisted process. | Keeping the ratio finding separate from the legality assessment creates a reviewable record of both the disparity evidence and the employer’s response. |
| 6 | Complete remediation and retain the audit, findings, defense materials, and deployment decision; do not deploy the affected AI-assisted process while the triggered review and remediation file remain unresolved. | Document first, then defend only on a supported record: a ratio finding alone is not a standalone determination of legality. |

## Frequently Asked Questions

**Which hiring stages should employers include when calculating AI-influenced selection rates?**

Employers should calculate and retain protected-group selection rates for every AI-touched stage, including early sourcing filters, recruiter prescreens, resume parsing, and other selection steps.

**How is the four-fifths selection-rate ratio calculated for each protected group?**

Divide each protected group’s selection rate by the highest selection rate observed among the groups.

**What does a selection-rate ratio below 0.80 indicate?**

It is evidence of adverse impact and should trigger investigation and documentation.

**Can an employer defend an AI hiring model if some stages meet the threshold but another stage falls below 0.80?**

No; a passing stage cannot offset a failing stage, and the model should not be defended as lawful merely because other stages meet the threshold.

**Does a four-fifths ratio finding by itself establish that a hiring practice is unlawful?**

No; a ratio below 0.80 is a defense-triggering signal requiring review, not a standalone determination that the hiring practice is unlawful.

**What should employers do after identifying an adverse-impact ratio in an AI-influenced hiring stage?**

They should investigate and document the finding at the stage where it occurred.

## Quick answers

| What should employers calculate and retain for each AI-influenced hiring stage? | Employers should calculate and retain protected-group selection rates for each stage, including early sourcing filters, recruiter prescreens, resume parsing, and other selection steps. |
| --- | --- |
| What event should follow a selection-rate ratio below the adverse-impact review threshold? | A ratio below the adverse-impact review threshold should trigger investigation and documentation. |
| Can a passing stage offset a failing stage in an AI-influenced hiring process? | No, a model should not be defended as lawful merely because other stages meet the threshold when any AI-influenced stage has a ratio below the adverse-impact review threshold. |
| What does a ratio below the adverse-impact review threshold establish? | It is a defense-triggering signal requiring review, not a standalone determination that the hiring practice is unlawful. |
| How is the four-fifths selection-rate ratio calculated for each protected group? | Divide the protected group’s selection rate by the highest selection rate observed among the groups. |

Also worth reading: **Which AI Hiring Tools Need a Bias Audit: NYC Local Law 144 Rules for 2026**: [Which AI Hiring Tools Need](https://ailaborbrain.com/blog/which-ai-hiring-tools-need-a-bias-audit-nyc-local-law-144-rules-for-2026.php) · **EEOC 2026 Bias Audits: Per-Hire Cost Up 30% to $52**: [EEOC 2026 Bias Audits: Per-Hire](https://ailaborbrain.com/blog/eeoc-2026-bias-audits-per-hire-cost-up-30-to-52.php) · **The 80% Rule: Why Vendor Bias Badges Won't Shield Employers**: [80% Rule: Why Vendor Bias](https://ailaborbrain.com/blog/the-80-rule-why-vendor-bias-badges-wont-shield-employers.php)

### Related reading

- [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)
- [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)
- [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)
- [7 Effective Steps to Document Workplace Harassment for HR Action A Data-Driven Approach](https://ailaborbrain.com/blog/7_effective_steps_to_document_workplace_harassment_for_hr_ac.php)
- [AI-Powered Document Verification How Neural Networks Streamline Name Changes at NJ DMV in 2025](https://ailaborbrain.com/blog/ai_powered_document_verification_how_neural_networks_streaml.php)

### Latest

- [Job exposure vs skills gap: Occupational Information Network (O*NET) or...](https://ailaborbrain.com/blog/job-exposure-vs-skills-gap-occupational-information-network-onet-or-postings-for-2026.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)
- [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)

Canonical: https://ailaborbrain.com/blog/ai-hiring-bias-audits-nycs-four-fifths-ruledocument-or-defend.php
Markdown: https://ailaborbrain.com/blog/ai-hiring-bias-audits-nycs-four-fifths-ruledocument-or-defend.php/index.md