What is adverse impact, and how do you measure it?
Adverse impact is a substantially different rate of selection that disadvantages a race, sex or ethnic group. You measure it by dividing each group's selection rate by the highest group's rate. Under the federal four-fifths rule, a ratio below 80 percent is generally treated as evidence of adverse impact rather than proof of it.
What adverse impact means
Adverse impact The federal Uniform Guidelines define it as "a substantially different rate of selection in hiring, promotion, or other employment decision which works to the disadvantage of members of a race, sex, or ethnic group" (29 CFR 1607.16).
A neutral-looking step screens one group out at a much lower rate than another. Nobody intended it. That definition of adverse impact turns on outcome, and the same regulation counts an informal interview or an unscored application form as a selection procedure.
How to calculate adverse impact in 4 steps
- Count applicants and selections in each group Pick one decision point and one group definition, then count how many people entered and how many were selected. Keep race, sex and ethnic group separate, as 29 CFR 1607.4(A) expects you to. Group A: 200 applicants, 100 selected. Group B: 100 applicants, 30 selected.
- Turn each count into a selection rate Divide selections by applicants for each group. The Uniform Guidelines define the selection rate as the proportion of applicants or candidates who are hired, promoted or otherwise selected. Group A: 100 of 200 is 50 percent. Group B: 30 of 100 is 30 percent.
- Find the highest rate Whichever group was selected at the highest rate becomes the benchmark. Every other group is compared against it, not against an average and not against the population. Group A at 50 percent is the highest rate here.
- Divide each rate by the highest to get the impact ratio That ratio is the number people mean when they say adverse impact ratio. Compare it to 80 percent. Anything below the line is the signal that starts an adverse impact analysis, not the conclusion of one. 30 divided by 50 is 0.6, so the impact ratio is 60 percent. Below 80.
A worked four-fifths rule example
| Group | Applicants | Selected | Selection rate | Impact ratio |
|---|---|---|---|---|
| Group A | 200 | 100 | 50 percent | Highest rate, so 100 percent |
| Group B | 150 | 66 | 44 percent | 88 percent, above the line |
| Group C | 100 | 30 | 30 percent | 60 percent, below the line |
| Group D | 80 | 38 | 47.5 percent | 95 percent, above the line |
The rule in one line
29 CFR 1607.4(D): a selection rate for any group "which is less than four-fifths (4/5) (or eighty percent) of the rate for the group with the highest rate will generally be regarded by the Federal enforcement agencies as evidence of adverse impact".
Where the four-fifths rule breaks down
It is unreliable on small numbers. The EEOC's Q&A puts it plainly: enforcement action is generally inappropriate where "the number of persons and the difference in selection rates are so small that the selection of one different person for one job would shift the result". The regulation adds that larger differences may not be adverse impact where they rest on small numbers and are not statistically significant.
It is also not a safe harbour. The EEOC calls it a rule of thumb that "is not intended to resolve the ultimate question of unlawful discrimination", and says it "merely establishes a numerical basis for drawing an initial inference". Where numbers are large, a gap under 20 percent can still be adverse impact if it is significant in statistical and practical terms.
Where adverse impact hides in a hiring funnel
- Where the job advert was placed, and who it was targeted at
- Knockout questions on the application form
- Resume screening, whether a person or a model does it
- Any test, assessment or work sample
- The informal phone screen, which the regulation still counts as a selection procedure
- The structured interview and the rubric behind it
- Background and credit checks late in the process
- The final offer decision, where most teams start and stop measuring
Adverse impact vs disparate treatment
The meaning of adverse impact sits in the outcome, not the intent. Adverse impact means a neutral practice produced a disproportionately negative effect on a protected group, which the EEOC treats as unlawful where the practice is not job related and necessary to the operation of the business. Disparate treatment is the opposite shape: someone is treated differently because of a protected characteristic, and intent is the whole question.
Title VII sets out how the impact argument runs. The complaining party has to show that a particular employment practice causes a disparate impact. The employer can then defend it as "job related for the position in question and consistent with business necessity". Note that "adverse impact" and "disparate impact" describe the same idea; the first is the Guidelines' word, the second is the statute's.
What to do when a ratio comes back under 80 percent
- Check the sample before you react Count how many people were actually selected. If one different decision would have flipped the ratio, you have noise, not a finding. Note it, widen the window, and look again.
- Find the step that produced it Recalculate stage by stage across the total selection process. A failing ratio at the offer usually traces back to a screen much earlier, and the fix belongs where the drop happened.
- Ask whether the procedure is job related Under 29 CFR 1607.3(A), a procedure with adverse impact is treated as discriminatory unless it has been validated. Write down what evidence links this step to performance on the job.
- Look for an equally valid alternative Where two procedures serve the same purpose and are substantially equally valid, the Guidelines say to use the one shown to have the lesser adverse impact. That is the remedy the regulation actually names.
- Take employment counsel before changing a standard Selection standards, record keeping and any adjustment to them carry legal consequences. This page is general information. Your counsel and your own data decide what happens next.
Fix the standard before anyone is interviewed
An example evaluation template: named criteria for one role, each weighted, with the total required to reach 100 before it can be saved.
The cheapest way to make an interview step defensible is to decide what you are measuring before you meet anyone. On The Cognitive, the evaluation criteria and weights are fixed per role and apply to everyone in it, while the questions stay live and adapt to each answer.
AI hiring law at a glance
| Rule | Who it covers | What it requires | Where to check |
|---|---|---|---|
| NYC Local Law 144 | Employers and agencies using an automated employment decision tool for a job in the city | A bias audit less than a year old before use, a published summary of results, and notice to NYC candidates 10 business days ahead | nyc.gov, the DCWP automated employment decision tools page |
| EU AI Act | Recruitment and selection systems, listed as high risk in Annex III point 4(a) | High risk means permitted with obligations, not banned: risk management, data governance, documentation, logging, transparency, human oversight | artificialintelligenceact.eu, and our EU AI Act page |
| Colorado SB24-205 | Deployers of high risk AI systems making consequential decisions, including employment | Reasonable care against algorithmic discrimination: a risk policy, impact assessments, annual review, consumer notice and an appeal | leg.colorado.gov. SB25B-004 pushed the compliance date to 30 June 2026, so check there for later changes |
| Federal Title VII | Every covered US employer, whatever software is or is not involved | A practice causing disparate impact must be job related and consistent with business necessity | eeoc.gov, plus the Uniform Guidelines at 29 CFR 1607 |
NYC Local Law 144 and automated employment decision tools
New York City Local Law 144 took effect on 1 January 2023, with enforcement from 5 July 2023. A tool is in scope when three things are all true: it uses machine learning, statistical modelling, data analytics or artificial intelligence; it helps make employment decisions; and it substantially assists or replaces discretionary decision making. The rule defines that last phrase narrowly and usefully.
Under the rule, a tool substantially assists or replaces a decision if you rely solely on its simplified output with no other factors, or weight that output more heavily than any other criterion, or use it to overrule conclusions drawn from other factors including human judgement. A simplified output is a score, tag, classification or ranking. Transcription and translation are expressly excluded.
The duties are practical. The bias audit happens before use and may be no more than a year old. The summary of results goes on the employment section of your website and stays up 6 months after last use. Candidates living in the city get notice 10 business days ahead. The employer carries this, not the vendor, and the law requires the audit rather than any result.
Questions to ask any AI hiring vendor
- Can we export selection or scoring rates by category, for our own audit?
- Is the output the sole factor, the heaviest factor, or able to overrule a person?
- Has this tool had a bias audit in the last 12 months?
- Did that audit use historical data or test data, and why?
- Did the auditor help build or sell the tool, work for either side, or hold a financial interest?
- Is any demographic data imputed or inferred rather than reported?
- What is the data retention policy, and what data is collected?
- Who reviews the output before a candidate is rejected?
Every decision keeps a person and a record behind it
An example scorecard: each criterion scored out of 5, an overall score out of 100 weighted from the role's evaluation criteria.
An audit needs outcome data, and a defence needs to show a human decided. On The Cognitive, each interview produces a recording, a transcript and a scorecard, and a recruiter still approves, rejects or holds the candidate. Pipeline stage history keeps the record of who moved where.
Keep going
The EU AI Act and hiring · Bias and fairness · What is a structured interview? · Free rubric generator
Frequently asked questions
What is adverse impact?
The Uniform Guidelines answer what is an adverse impact in hiring: a substantially different rate of selection that works to the disadvantage of a race, sex or ethnic group. Intent is irrelevant. The regulation reaches any measure used as a basis for an employment decision, including informal interviews and unscored application forms.
What is the four fifths rule?
The four fifths rule, sometimes written as the four fifth rule, comes from 29 CFR 1607.4(D). A selection rate for any group that is less than four-fifths, or 80 percent, of the rate for the highest group is generally regarded by federal enforcement agencies as evidence of adverse impact. A rate above it generally is not. A rule of thumb, not a legal standard.
How do you calculate adverse impact?
Four steps. Count applicants and selections for each group. Divide to get each group's selection rate. Find the highest rate. Divide every other rate by that highest rate. The result is the impact ratio, which you compare to 80 percent. Calculating adverse impact one funnel stage at a time tells you far more than doing it once at the offer.
What is an adverse impact ratio?
It is one group's selection rate divided by the selection rate of the most selected group, expressed as a percentage. New York City's AEDT rule writes the same formula and adds a second version based on scoring rates, where the scoring rate is the share of a group scoring above the sample's median score.
Can you give an adverse impact example?
Take two groups at one screening step. Group A: 200 applicants, 100 selected, a 50 percent selection rate. Group C: 100 applicants, 30 selected, a 30 percent rate. Group A has the highest rate, so the impact ratio for Group C is 30 divided by 50, or 60 percent. That is below 80, so it is evidence worth investigating.
Does the four fifths rule work when only a few people are hired?
Poorly. The EEOC says enforcement action is generally inappropriate where the number of persons and the difference in selection rates are so small that selecting one different person for one job would shift the result. The regulation says larger differences may not indicate adverse impact when they rest on small numbers and are not statistically significant. Pool more data first.
Does passing the four fifths rule mean we are legally safe?
No. The EEOC says the 4/5ths rule of thumb speaks only to adverse impact and is not intended to resolve the ultimate question of unlawful discrimination. Where numbers are large, a gap of less than 20 percent can still be adverse impact if it is significant in statistical and practical terms. Passing means no obvious signal, not a defensible practice.
What is an automated employment decision tool?
Under NYC Local Law 144 it is a computer-based tool that uses machine learning, statistical modelling, data analytics or artificial intelligence, helps employers make employment decisions, and substantially assists or replaces discretionary decision making. That last test means relying solely on its score, weighting the score above every other criterion, or using it to overrule human conclusions.
What does NYC Local Law 144 require?
A bias audit by an independent auditor before use and no more than a year old, a published summary of results on the employment section of your website, and notice to candidates who live in New York City at least 10 business days before the tool is used. The audit covers sex, race or ethnicity and intersectional categories. The employer, not the vendor, must make sure it happened.
Does Local Law 144 have a quarterly reporting requirement?
No. Nothing in the law, the rule or the DCWP guidance imposes quarterly reporting. The bias audit must be less than 1 year old, and the published summary of results has to stay posted for at least 6 months after the last use of the tool. Check the DCWP automated employment decision tools page on nyc.gov for the current position.
Explore: The EU AI Act and hiring · Bias and fairness in AI interviews · What is a structured interview? · AI bias in hiring, a practical guide