The Four-Fifths Rule in a Bias Audit, Explained

Reading an Impact Ratio Without Panicking

Your AEDT vendor just sent over the annual bias audit summary, the one you're required to have posted before the tool runs on another NYC applicant. You open the PDF. Selection rate for one group: 42%. Selection rate for another: 31%. A column labeled "impact ratio": 0.74. Nobody attached a note explaining what that number is supposed to mean, and the vendor's account rep is out until Monday.

This is the moment a lot of HR directors either panic and pull the tool offline, or shrug and file the PDF without reading it at all. Neither is right. That 0.74 is doing a specific, narrow job — it's flagging something worth a second look, not handing down a verdict on your hiring process.

By the end of this piece, you'll be able to read that impact ratio line the way it's meant to be read: as a screening signal built into the mechanics of a bias audit, with a known origin, a known limitation, and a known next step.

What the Four-Fifths Rule in a Bias Audit Actually Measures

The four-fifths rule compares selection rates between groups. If one group is selected (hired, advanced, scored favorably) at a rate that's less than 80% of the rate for the highest-selected group, that comparison is flagged as a possible sign of adverse impact.

That's it. It doesn't measure intent. It doesn't measure whether the AEDT's underlying model is "biased" in some technical, mathematical sense. It measures one thing: whether the outcomes for two groups, side by side, land close enough together that the gap probably isn't the first place to look for a problem — or far enough apart that it is.

This is why the four-fifths rule shows up as a standard line in almost every published Local Law 144 bias audit summary. It's a common, portable reference point that auditors, regulators, and employers can all read the same way, without each auditor inventing its own threshold.

Where the 80% Reference Point Comes From

The four-fifths threshold isn't a Local Law 144 invention. It comes from the EEOC's Uniform Guidelines on Employee Selection Procedures, which predate AI hiring tools entirely.

A selection rate for any race, sex, or ethnic group that is less than four-fifths (or 80%) of the rate for the group with the highest selection rate will generally be regarded as evidence of adverse impact.

LL144's bias audit requirement borrowed this pre-existing employment-law reference point rather than writing a new statistical standard from scratch (via Assessment Systems, 2024). That's worth knowing because it means the four-fifths rule bias audit connection isn't NYC-specific — it's a decades-old EEOC guideline applied to a new category of tool: the automated employment decision tool as defined under Local Law 144, meaning any computational process derived from machine learning, statistical modeling, or AI that issues a score, classification, or recommendation used to substantially assist or replace hiring or promotion decisions (Perkins Coie, 2023).

How to Calculate an Impact Ratio: A Worked Example

The math itself is simple enough to do in the margin of the audit summary.

Step 1 — find the selection rate for each group. Say a tool advances 80 out of 100 applicants in Group A (80%) and 60 out of 100 applicants in Group B (60%).

Step 2 — divide the lower rate by the higher rate. 60% ÷ 80% = 0.75.

Step 3 — compare the result to 0.80. 0.75 is below 0.80, so this comparison would be flagged under the four-fifths rule.

That's the whole calculation. A published LL144 audit summary is required to show selection rates (and, depending on the AEDT's function, scoring rates) broken out by race/ethnicity and sex categories, along with the resulting impact ratios, plus the date the AEDT was made available for use (Crowell & Moring LLP, 2023; Epstein Becker Green, 2023). If you can find the selection-rate columns, you can reconstruct the ratio yourself rather than taking the summary's framing at face value.

What a Ratio Below 0.80 Means — and What It Doesn't

A ratio under 0.80 means the gap between two groups' outcomes has crossed a threshold that's conventionally treated as worth investigating. It does not mean:

• The tool is illegal to keep using.
• The employer has violated anti-discrimination law.
• The vendor has done something wrong.
• The audit has "failed."

Small sample sizes distort this ratio badly. If only 12 candidates from a given group applied in the audit period, one or two outcomes can swing the ratio well below 0.80 without reflecting any real pattern. A published bias audit summary is a snapshot of one period's numbers, not a certification that the tool is safe (or unsafe) in every deployment context going forward.

This is where it's worth being direct about what these workbooks do and don't do: nothing here is legal advice, and nothing in a downloadable compliance workbook performs, certifies, or signs a bias audit. If a ratio in your vendor's summary looks concerning, the next step is a conversation with the independent auditor who ran the audit, or with counsel — not a guess based on one number in isolation.

Reading the Four-Fifths Line in a Published LL144 Bias Audit Summary

In practice, this is harder than it should be — not because the math is complicated, but because a lot of published summaries are thin. A 2024 academic study of 391 employers subject to LL144 found that only 18 had posted an audit report and only 13 had posted the required transparency notice at all (ACM FAccT, Wright & Muenster et al., 2024). If you're trying to benchmark your own vendor's summary against "what a normal one looks like," there often isn't much of a normal to compare against.

A separate 2025 study of published LL144 audits found that many under-report disparities — missing demographic data, opaque aggregation across job categories, and metrics that don't reflect how the tool is actually deployed can all make an impact ratio look cleaner than the underlying reality (ACM FAccT, 2025). None of that means the four-fifths line is useless. It means the ratio is only as good as the selection-rate data underneath it, and reading it well means checking whether that data is actually granular and complete — not just glancing at the final number.

If you want a structured way to walk through a summary line by line, that's a narrower read than this piece; see how to read a bias audit summary for the fuller breakdown, and selection rate vs. scoring rate if your AEDT reports scores rather than binary selections.

Where the Four-Fifths Rule Breaks Down in Practice

A few honest limitations worth flagging before you rely on this rule too heavily:

It's a rule of thumb, not a legal standard. Courts and regulators can and do look past a passing four-fifths ratio when other evidence points to disparate impact, and a failing ratio doesn't automatically establish a violation either.

It says nothing about causation. A low ratio tells you two groups had different outcomes. It doesn't tell you why — that requires digging into the AEDT's inputs, the applicant pool, and the job requirements it's screening against.

It's sensitive to how groups are defined and aggregated. Combine or split demographic categories differently and the same underlying data can produce a different ratio.

If your organization is trying to understand the mechanics one level up from the four-fifths line — what LL144 requires overall, who has to comply, and what the filing calendar looks like — that context lives in the Local Law 144 compliance guide and the overview of what a bias audit actually covers. For the deeper mechanics of the impact ratio calculation itself, including how scoring-based AEDTs are handled differently from pass/fail ones, see impact ratio under Local Law 144.

Your First Action Item

You don't need to become a statistician to read a bias audit summary competently. You need a repeatable way to pull the selection-rate numbers, run the ratio, and know what a result under 0.80 should — and shouldn't — trigger on your end.

That's what the Four-Fifths Rule & Impact-Ratio Reading Guide is built for: a plain-language PDF walkthrough plus a companion calculator worksheet, so the next time a vendor's summary lands in your inbox, you're checking the math yourself instead of taking the final line on faith. Download it, run your current vendor's most recent published summary through it, and note anywhere the numbers don't add up before your next renewal conversation.