Naledi runs talent acquisition for a 600-person fintech with a 5-person recruiting team, and she learned the hard way that "we treat everyone the same" is not a bias program. After a quarterly review, she pulled callback rates by demographic for one high-volume engineering req and found the team had interviewed 40 percent of male applicants but only 28 percent of female applicants. Nobody had decided to favor men. The bias lived in language, in requirements, and in an AI resume screener that had quietly learned the patterns of past hires. This lesson is the toolkit Naledi built afterward: how each major bias category shows up across sourcing, screening, interviewing, and AI tools, what the law says, and how to measure whether your process is actually fair.

Why Bias Categories Behave Differently

Bias does not operate uniformly. A woman faces different bias than a man. An older worker faces different bias than a younger one. A candidate with a disability faces different bias than a non-disabled candidate. A person of color faces different bias than a white person. Someone from a working-class background faces different bias than someone from privilege. Gender bias, age bias, disability bias, racial bias, and socioeconomic bias each travel through recruiting along different routes and trip different laws, so a generic "be fair" intervention misses the specific mechanisms. Naledi treats each category as its own detection problem with its own legal exposure.

This is not abstract learning. It is learning to see patterns in your actual recruiting. When you understand that women in technical roles face specific skepticism about technical ability, you can watch for it in your own interview feedback. When you understand that candidates of color face accent discrimination, you can notice whether you are questioning communication competence based on accent. When you understand that older workers face assumptions about technology aptitude, you can catch yourself making them. The goal is not to memorize manifestations; it is pattern recognition sharp enough that you notice bias while it operates, because that recognition is the precondition for changing the process that produced it.

The legal map matters because it tells you where the financial and reputational risk concentrates. Title VII of the Civil Rights Act covers race, color, sex, religion, and national origin. The Age Discrimination in Employment Act (ADEA) protects workers aged 40 and over. The Americans with Disabilities Act (ADA) covers disability and governs accommodation. The EEOC Uniform Guidelines on Employee Selection Procedures define how the government measures whether a selection process has adverse impact, using the four-fifths rule.

Two further points complete the map. If you use an automated employment decision tool to screen or rank candidates for a New York City role, NYC Local Law 144 requires an independent bias audit and candidate notice. Socioeconomic status is the outlier: it is not a protected class federally, yet its proxies (school prestige, unpaid internships, relocation ability) often correlate with race and sex, which means socioeconomic filters can produce protected-class adverse impact through the back door.

Measuring Adverse Impact: The Four-Fifths Rule

Before diving into categories, Naledi grounds her team in the one number the EEOC actually uses. The four-fifths (80 percent) rule compares selection rates across groups. You calculate the selection rate for each group, find the group with the highest rate, and divide every other group's rate by it. If any ratio falls below 80 percent, that is evidence of adverse impact and a signal to investigate the step that caused it.

Here is the worked example from Naledi's engineering req. The resume screener advanced candidates from the applicant pool to a phone screen. Of 250 male applicants, 100 advanced, a selection rate of 40 percent. Of 150 female applicants, 42 advanced, a selection rate of 28 percent. Men are the highest-selected group, so the comparison is 28 percent divided by 40 percent, which equals 0.70, or 70 percent. That is below the 80 percent threshold, so the screening step shows adverse impact against women under Title VII.

The ratio does not prove illegal discrimination by itself; an employer can defend a selection procedure that is job-related and consistent with business necessity. But a 70 percent ratio shifts the burden. Naledi has to either show the screen measures something the job genuinely requires, or fix it. When she audited the screener, she found it had down-weighted resumes with employment gaps and resumes lacking certain "elite" employer names, both of which correlated with the women in her pool who had taken caregiving leave. The rule did exactly its job: it pointed at the step, and the step pointed at the mechanism.

Run this calculation at every funnel stage, not just the hire. Sourcing, screening, interview-to-offer, and offer-to-accept each have their own selection rates, and bias can hide at any one while the overall hire rate looks balanced. When a group's rate is lower, the follow-up question is what job-relevant explanation appears in the documentation.

How Gender Bias Manifests

Gender bias enters Naledi's funnel through interpretation more than exclusion, and it takes five recognizable forms. The first is role-based bias. Women in technical roles face skepticism about technical ability that men do not. When a woman describes her technical background, evaluators may question its depth or ask for more proof; when a man describes the same background, evaluators assume competence. This operates even when her qualifications are identical or stronger.

The second is leadership bias, where women in leadership roles get judged on different dimensions altogether. A man is evaluated on problem-solving ability and strategic thinking. A woman in the same role is evaluated on interpersonal skills and communication style. When both are equally strong technically, the man is described as a "strong leader" and the woman as a "good communicator." The same underlying quality, good interpersonal communication, is interpreted differently depending on who displays it, and only one of those descriptions reads as promotable.

The third is the motherhood penalty. When women disclose motherhood or caregiving responsibilities, assumptions activate about commitment and availability that do not activate for men. A man who says "I have family commitments" does not trigger assumptions about job focus. A woman saying the same thing triggers assumptions about willingness to travel, work overtime, or relocate. The disclosure is identical; the inference is not.

The fourth is interpretation bias, where the same behavior is read differently by gender. A man is "ambitious" and "driven"; a woman showing equal ambition is "pushy" or "aggressive." A man is "decisive"; a woman is "bossy." A man "speaks his mind"; a woman "does not play well with others." A man is "negotiating for fair value"; a woman is "difficult about compensation." No single one of these decides an outcome, and that is precisely the problem: they accumulate quietly across an evaluation until the summary reads differently for two people who behaved the same way.

The fifth is communication style bias, and it produces a bind with no exit. Direct communication is valued in men and penalized in women. Collaborative communication is valued in women but read as weak leadership in men. A woman is therefore penalized whichever style she chooses, which means the fix cannot be advice to the candidate; it has to be a change in how the evaluation is conducted.

Detecting gender bias takes specific attention rather than good intentions. Do you evaluate men and women on the same dimensions? When two candidates have similar qualifications, does gender shift the interpretation? Are communication styles read the same way? Do mothers and fathers get different questions about flexibility, travel, and commitment? And do you hire men and women at the same rates, and if not, what job-relevant explanation appears in the documentation?

Naledi reads feedback in pairs, taking two debriefs for similarly qualified candidates of different genders and checking whether the same behavior earned the same words. She uses an AI assistant such as Anthropic's Claude to flag asymmetric language across a batch of debriefs, surfacing where one candidate gets strength words and another caution words for equivalent evidence. The AI does not decide; it makes the pattern visible faster than manual reading.

How Age Bias Manifests

Age bias is governed by the ADEA, which protects applicants 40 and over, and it surfaces through proxies rather than explicit age questions. The first form is technology bias: older workers are assumed to struggle with new technology or to be unwilling to learn it, while younger workers are assumed to be "digital natives." These assumptions persist despite evidence that learning ability does not correlate with age and that older workers adapt to technology effectively.

The second is energy and stamina bias. Younger workers are assumed to be hungrier and willing to work longer hours; older workers are assumed to lack energy or stamina. It shows up in questions like "are you comfortable with our fast-paced startup culture?" asked to older candidates and not younger ones: a differential question rather than a differential answer.

The third is commitment bias, and it runs in both directions. Older workers are assumed to be close to retirement and likely to leave soon. Younger workers are assumed to be job-hoppers who will leave for a slightly better offer. So older candidates face a "you will leave for retirement" assumption and younger candidates face a "you will leave for money" assumption, both of them commitment bias, neither of them evidence.

The fourth is salary bias. Older workers are assumed to be expensive and to require high compensation; younger workers are assumed to be satisfied with less. This affects offer decisions directly, and two candidates with equal qualifications can receive different offers on the strength of an age-driven assumption about what each will accept.

The fifth is cultural fit bias, which here means age similarity. People tend to perceive cultural fit as being close in age to the existing team. An all-young team reads younger candidates as good fits and older ones as potential misfits; an all-older team does the reverse. Either way the effect is homogeneous hiring that compounds over time.

Detection here is mostly about language and questions. Watch for graduation-year fields, "digital native" phrasing, "high-energy startup culture" framed as a question only older candidates get, and salary assumptions that older workers will be too expensive. A job posting that says "recent graduate" or "0 to 3 years only" is a near-automatic ADEA flag because it screens for youth. Ask whether assumptions about technology, energy, or commitment appear in your feedback, and disaggregate hiring data by age. Naledi removed graduation dates from the fields her screener could see and ran the four-fifths rule on advance rates for the over-40 group; a ratio below 80 percent there is age adverse impact even when no one asked anyone's age.

How Disability Bias Manifests

Disability bias under the ADA runs through five mechanisms. The first is capability assumptions: when a disability is disclosed or apparent, evaluators often assume limitations on job performance that may not exist. A blind candidate applying for a programming role might face questions about how they will write code, despite screen readers and adapted interfaces enabling full participation.

The second is accommodation anxiety. Evaluators assume disclosure means extensive, burdensome accommodation needs, overestimating cost and underestimating what accommodations enable. In practice, screen readers, captioning, and flexible scheduling enable full participation at low or no expense. The rejection is driven by fear rather than actual job requirements.

The third is disclosure discrimination. Where a disability is not visible and not disclosed, evaluators may show no bias at all; when it is disclosed, bias activates. The predictable effect is pressure on candidates to conceal disabilities in order to be evaluated fairly, which is a poor outcome for everyone and a signal that the process, not the candidate, needs to change.

The fourth is ableist language and requirements. Job descriptions routinely carry requirements that exclude disabled people without anyone examining whether the requirement is necessary. "Must be able to work in an open office environment" excludes people with sensory sensitivities. "Requires a fast-paced environment" may exclude people who work at a different pace. "Must work standard business hours" excludes people who need flexible scheduling for medical reasons.

The fifth is inaccessible process bias, where the recruiting process itself excludes candidates before any evaluation happens. Phone-only screens exclude Deaf candidates. In-person-only interviews exclude people with mobility disabilities. Unstructured group interviews disadvantage candidates with anxiety conditions. These are barriers to entry, not assessments, and they remove qualified people before anyone has looked at their qualifications.

Detection questions follow the mechanisms. Do you automatically assume limitations based on a disability? Do you ask about medical history or personal accommodation needs? The ADA bars medical and disability inquiries before a conditional offer, so those questions during screening are themselves legal exposure. Do your job descriptions include requirements that might not be necessary? Is your process accessible? And when candidates disclose disabilities, do they advance at the same rates as non-disabled candidates? Naledi audits every requirement against whether it is an essential function of the job, and offers candidates a standard accommodation option at scheduling without asking why.

How Racial and Ethnic Bias Manifests

Racial and ethnic bias under Title VII is the most documented and the most proxy-laden, and it takes six forms. The first is name-based callback bias: identical resumes receive different callback rates depending on whether the name is perceived as belonging to a particular racial or ethnic group. "Jamal Williams" gets fewer callbacks than "Jake Williams." "Fatima Khan" gets fewer callbacks than "Fiona Kelly." This happens at the application stage, before anyone has evaluated an actual qualification.

The second is accent discrimination. Candidates are penalized for accents that signal a non-native English speaker or non-U.S. origin. Even where communication is clear and professional, an accent triggers assumptions about competence, communication ability, or cultural fit that are not grounded in observed ability.

The third is educational institution bias. Certain institutions signal socioeconomic privilege and also correlate with race, so the absence of a "prestige school" triggers assumptions about qualification that take no account of structural inequality in educational access.

The fourth is location and background bias. Requirements such as "must have worked in a major tech hub" or "must have Bay Area experience" correlate with race and geography, filtering out people from underrepresented groups while sounding entirely neutral in a requisition.

The fifth is cultural fit bias. "Culture fit" evaluations frequently select for people similar to the existing team, so if that team is predominantly white, culture fit becomes code for preferring white candidates. The usable version of the concept is "can work effectively in our environment," not "is similar to the people already here."

The sixth is network bias. Recruiting through employee referrals, which is extremely common, perpetuates racial homogeneity because personal networks are often racially segregated. Referral bonus systems amplify the effect by paying for it.

Detection means looking at callback data by name, tracking outcomes through every recruiting stage rather than only at hire, examining how feedback differs for candidates of different races, auditing requirements for correlation with race, and analyzing whether culture fit evaluations perpetuate homogeneity. Naledi's fixes are structural. She removed names and photos from the screening view to neutralize name-based callback bias. She replaced "culture fit" in scorecards with specific, observable "ways of working" criteria. And she ran the four-fifths rule on callback rates by race the same way she had run it on the gender gap that started all of this; where a group's rate falls below four-fifths of the highest group's rate, the next step is to trace the mechanism rather than to assume intent, which is how a "Bay Area experience" requirement that nobody could justify came out of the job description.

How Socioeconomic Bias Manifests

Socioeconomic bias is often invisible precisely because socioeconomic status is not a protected class, so nobody is watching for it, and it is powerful anyway. The first form is requirement-based bias. Certain requirements correlate strongly with wealth: unpaid internships, the ability to relocate, access to particular educational institutions, and access to professional networks. These requirements do not measure job ability. They measure privilege.

The second is background assumptions. Candidates without elite school backgrounds face unwarranted assumptions about work ethic, ambition, or capability. A candidate from a non-target school or non-tech background is assumed to lack opportunity or sophistication, when what they lack is the credential the evaluator is used to seeing.

The third is culture fit bias again, in its class version. When culture fit includes values or interests that correlate with privilege, such as hiking, craft beer, or international travel, candidates from working-class backgrounds are filtered out by a social screen that has nothing to do with the work.

The fourth is communication style bias. Communication styles correlate with socioeconomic background, and candidates from privileged backgrounds tend to use styles that evaluators perceive as more professional or sophisticated. Candidates from working-class backgrounds may use different styles that read as less polished despite equal competence, and "polished" is doing the scoring.

The fifth is the treatment of gaps and unconventional paths. A candidate who worked while going to school, who has employment gaps caused by financial necessity, or who took a non-linear route into the field faces bias in how that history is read. The same facts that could be described as resourcefulness get logged as a red flag.

Detection starts with your requirements: which correlate with privilege rather than ability? Then with how you interpret gaps and unconventional paths, as risk or as evidence of resourcefulness. Then with culture fit: does participating in your culture require privilege? The legal hook is that these proxies frequently correlate with race and sex, so a socioeconomic filter can generate protected-class adverse impact that the four-fifths rule will catch even though "class" never appears in the data.

Where AI Tools Concentrate All of It

AI tools concentrate every risk above because they learn from history. A resume screener trained on a company's past hires inherits the demographics of those hires; if past engineers skewed male and came from three universities, the model learns those signals and penalizes everyone else, exactly what happened in Naledi's opening example. This is why NYC Local Law 144 requires an independent bias audit of any automated employment decision tool used for hiring or promotion of NYC candidates, requires that the audit results be published, and requires that candidates be notified the tool is in use. Even outside New York, the EEOC has stated that an employer remains liable for adverse impact produced by a vendor's AI tool; "the algorithm did it" is not a defense.

Naledi's AI governance has three rules. First, audit before deploy and on a schedule after: run the four-fifths rule on the tool's outputs by every protected group before it touches a real candidate, then quarterly. Second, constrain what the tool sees and says: she instructs assistants such as Claude or ChatGPT to evaluate only against job-relevant criteria and to never infer or use protected characteristics, and she keeps names, ages, and photos out of the screening view entirely. Third, keep a human accountable for every advance and reject decision, with the AI producing evidence-based recommendations rather than final rankings. The tool accelerates the work; the recruiter owns the outcome.

Intersectional Bias and Detection

These biases do not operate independently. A woman of color faces gender and racial bias simultaneously, and not additively but multiplicatively: she faces gender stereotypes about ambition and communication style and racial stereotypes about capability and cultural fit at the same time, and the combination is more powerful than either alone. The intersection produces stereotypes that neither category predicts on its own.

The pattern repeats across combinations. An older woman faces age and gender bias together; she is read as "too old" and "too demanding" where an older man is read as "experienced." An older person of color faces age, racial, and potentially ethnic bias at once. An immigrant with a disability faces racial or ethnic bias, accent discrimination, disability bias, and potentially gender bias simultaneously. The intersections compound rather than queue.

This is why single-dimension analysis is dangerous. Naledi's overall gender hire rate looked balanced, but when she broke gender down by race she found white men and white women hired at similar rates while women of color advanced at a markedly lower one. The four-fifths rule on the combined group surfaced what the single-axis view had hidden completely. The practical rule is to disaggregate: when you measure by gender, break it down by race; when you measure by age, break it down by gender; when you examine outcomes, look at gender and race combinations. Compute selection rates for intersecting groups, not just headline demographics, because that is where compounding bias lives and where the largest legal exposure sits.

Subtle Manifestations

Most of what this lesson describes shows up subtly rather than explicitly, and the subtle version is both harder to see and more common. The first channel is tone. Feedback on a woman reads "was assertive" while feedback on an equally assertive man reads "strong leadership." Same behavior, different tone, opposite valence. Nothing in either note is factually wrong, which is why neither gets challenged.

The second channel is how much explanation a candidate is required to supply. Credentials from certain groups need longer justification: a candidate from a non-target school has to explain their background more, and a candidate with an unconventional path has to convince the evaluator that the path still taught relevant skills. The burden of proof shifts without anyone deciding to shift it.

The third channel is different standards applied to the same behavior. Ambition in men is a strength; ambition in women is a concern. Directness in men is efficient; directness in women is abrasive. The fourth is different interpretations of identical information. A resume gap for a man reads "he took time to travel and reflect, good for growth." The same gap for a woman reads "she probably left to have kids, may not be committed." Same information, opposite conclusion.

Detecting any of this requires attention to tone, to the standards applied, and to how information is interpreted rather than only to what is decided. Do different candidates get explained the same way? Do you use the same language and tone for all of them? Reading feedback in matched pairs, as Naledi does, is the most reliable way to make the asymmetry visible, because the difference only shows up in comparison.

Anti-Patterns

The assumption of homogeneity. This is assuming all bias operates the same way and can be addressed with generic fairness interventions. It happens because one-size-fits-all feels more manageable than five separate detection problems. What goes wrong is that you miss the specific manifestations: a generic intervention may address some forms while leaving others untouched, and you will not know which is which. Gender bias in technical roles looks nothing like age bias in startup culture. Understand the specific forms bias takes by demographic, and target interventions at those manifestations rather than at bias in general.

The invisible bias. This is failing to recognize bias because it arrives through subtle interpretation rather than explicit statement. Nobody says "we will not hire women." Someone says "she seemed pushy." It happens because bias is covert and its subtle forms feel defensible, since each comment can be justified on its own terms. What goes wrong is that you never recognize it and therefore never address it, so subtle bias persists in a process staffed entirely by people with good intentions. The counter is comparison: notice tone, interpretation, and the standards applied to different candidates, and look for patterns in how you read the same information about different people.

The single-dimension analysis. This is analyzing hiring disparities by one demographic at a time, gender or race, without examining the intersections. The classic example is finding that you hire men and women at similar rates and concluding you have no gender bias, then discovering on a race breakdown that white men and white women are hired at similar rates while men and women of color are hired at different ones. It happens because intersectional analysis takes more work. What goes wrong is that you miss how biases compound for people with multiple underrepresented identities, which is where both the harm and the legal exposure are greatest. Break gender down by race, break age down by race and gender, and look for compounding effects rather than headline averages.

Practice

Each of these produces something concrete, and all five are worth doing with your legal or DEI partner rather than alone.

• Run a specific bias audit. For each characteristic covered here, gender, age, disability, race, and socioeconomic status, identify how that bias shows up in your recruiting. What questions might you ask differently? What assumptions might you make? Which requirements filter certain groups out before evaluation?
• Analyze your callbacks. Pull callback data by demographic. Do the rates differ, and for which groups? Where a group's rate is lower, what job-relevant reason explains the difference, and is that reason documented anywhere or only assumed?
• Audit your interpretations. Take three recent debriefs and read the feedback on each candidate. Does tone or interpretation differ by candidate demographic? Are the same behaviors described the same way? Look for "assertive" against "aggressive," "leader" against "bossy," "rigorous" against "difficult."
• Audit your requirements. List every requirement on a current open role and ask what each one actually measures. An unpaid internship measures family wealth. An Ivy League degree correlates with privilege and with race. Must work in-office correlates with ability and transportation. Must work overtime correlates with caregiving status. Which of these exclude qualified candidates without measuring job ability?
• Analyze intersectionally. Compare outcomes for women and men overall, then break that down by race. Compare age groups overall, then break that down by gender. Where do disparities compound, and would your single-axis reporting have shown them?

Reflection

• Which form of demographic bias do you most recognize in your own recruiting, and what specific manifestations have you noticed?
• Have you seen interview feedback that might reflect subtle bias through different tone, different standards, or different interpretation for different candidates?
• Which of your requirements might correlate with demographic characteristics, and are they genuinely necessary for the job?
• If you ran an intersectional analysis of your hiring today, where do you suspect you would find compounding effects?
• What would change in your recruiting if you specifically watched for each type of bias described here?

Glossary

• Accent discrimination. Penalizing a non-native or non-standard accent despite clear communication, based on an unconscious association between accent and competence or fit.
• Adverse impact. A neutral-looking selection step that produces substantially lower selection rates for a protected group, measured under the EEOC Uniform Guidelines.
• Age bias. Assumptions about technology ability, energy, commitment, or cultural fit based on age rather than on individual capability or interest.
• Ableist language. Language or requirements designed for non-disabled people that inadvertently exclude disabled people without anyone examining whether the requirement is necessary.
• Cultural fit bias. Using "culture fit" as code for similarity to the existing team, which perpetuates homogeneity and filters out people who are different.
• Disability bias. Assumptions about capability limitations, accommodation needs, or the meaning of disclosure based on disability rather than on individual ability or actual job requirements.
• Gender bias. Assumptions about roles, caregiving, communication, or ambition based on gender rather than on individual capability or performance.
• Intersectional bias. The compounding effect of multiple biases on people with multiple underrepresented identities, more powerful than those biases operating independently.
• Motherhood penalty. Assumptions about commitment, availability, or willingness to travel or work overtime based on motherhood status, not applied to fatherhood.
• Name-based callback bias. Differential response rates to identical applications based on whether the name is perceived as belonging to a particular racial or ethnic group.
• Racial and ethnic bias. Assumptions based on perceived race or ethnicity, appearing through name discrimination, accent discrimination, educational institution, requirements, or culture fit.
• Socioeconomic bias. Requirements and assumptions correlating with wealth and privilege, such as unpaid internships, ability to relocate, and prestige schools, which filter for class background.
• Four-fifths rule. The EEOC screen under the Uniform Guidelines: divide each group's selection rate by the highest group's rate, and treat any ratio below 80 percent as evidence of adverse impact warranting investigation.

Related Lessons

• Sources of Bias: Data, Algorithms, Humans, and Systemic Factors explains where these manifestations originate, which determines whether a fix belongs in the data, the tool, or the process.
• Visible and Hidden Bias: What Stands Out and What's Subtle develops the subtle manifestations section in more depth, particularly the tone and interpretation channels.
• Fairness Interventions: Blind Reviews, Structured Processes, Diverse Panels covers the lawful remedies referenced throughout this lesson in practical detail.
• Fairness Metrics: Defining and Measuring Bias in Outcomes goes deeper on the measurement behind the four-fifths calculation, including small samples and intersectional cuts.
• Accessibility and Inclusion: Ensuring AI Processes Work for All handles the accommodation and process-accessibility obligations that sit alongside the disability bias analysis here.
• Hands-On Project: Audit a Recruiting Workflow for Bias is where this toolkit gets applied end to end against a real workflow of your own.

Closing

Specific bias forms require specific detection methods, which is the single practical claim of this lesson. Understanding what gender bias looks like in technical roles lets you watch for it. Understanding what racial bias looks like in callback rates and feedback lets you notice it. Understanding what age bias looks like in questions about technology and energy lets you catch yourself asking one. That recognition is the moment change becomes possible.

Specificity moves you from abstract awareness, "bias exists," to concrete recognition, "I just did that." Once you can see your own biases operating in real time, you can interrupt them, and the questions to have ready are short: why am I interpreting this behavior differently for different candidates? Why do I have this requirement? What am I assuming about this person based on their identity? Naledi's team did not fix its 70 percent ratio by trying harder to be fair. They named the mechanism, removed the requirement that produced it, constrained what the tool could see, and kept measuring at every stage including the intersections.

Key Takeaways

• Bias takes specific forms by demographic, and each has its own law. Gender bias in technical roles looks different from gender bias in administrative roles, and age bias in a startup looks different from age bias in a mature company. Title VII covers race and sex, the ADEA covers age 40 and over, the ADA covers disability, the EEOC Uniform Guidelines define adverse impact, and NYC Local Law 144 requires a bias audit for automated hiring tools. Match your detection to the specific category, not a generic fairness slogan.
• The four-fifths rule turns fairness into a number you can run. Divide each group's selection rate by the highest group's rate; a ratio under 80 percent flags adverse impact at that step. Run it at sourcing, screening, interview, and offer, not just the final hire.
• Subtle manifestations are hardest to recognize and most pervasive, and same behavior, different interpretation is the mechanism. Explicit bias is obvious and legally risky; subtle bias, such as "we are looking for hunger and energy," is harder to see and far more common. When ambition is read as "driven" for men and "aggressive" for women, that interpretive difference is where bias operates. Watch tone, interpretation, and the standards applied to the same behavior, not only decisions.
• Requirements often hide demographic filters. An Ivy League requirement filters for race and class. Must relocate filters for caregiving status. Must have worked in a tech hub filters for geography and privilege. Audit every requirement against essential job function, asking what it actually measures rather than what it appears to measure.
• Socioeconomic bias is unprotected but still actionable. Its proxies correlate with race and sex, so class filters produce protected-class adverse impact that the four-fifths rule will catch even when class never appears in the data.
• AI tools concentrate risk because they learn from history. A screener trained on past hires reproduces past demographics. Audit tools before deploy and on a schedule, keep protected attributes out of the view, instruct assistants like Claude or ChatGPT to use only job-relevant criteria, and keep a human accountable for every decision.
• Intersectional bias is multiplicative, not additive. A woman of color does not simply face gender bias plus racial bias; she faces a specific intersection that is often more powerful than either alone. Disaggregate by intersecting groups, because balanced headline numbers can hide a sharply lower advance rate for people with multiple underrepresented identities, which is also where the largest legal exposure concentrates.

Frequently Asked Questions

Our four-fifths ratio is below 0.80. Does that mean we are discriminating? Not by itself. The ratio is evidence of adverse impact and a signal to investigate the step that produced it, and an employer can defend a selection procedure that is job-related and consistent with business necessity. What it does is shift the burden: you now need to show the step measures something the job genuinely requires, or fix it. Naledi's 0.70 pointed at her screening step, and the screening step pointed at employment gaps and elite employer names, neither of which was defensible as job-related.

Socioeconomic status is not protected. Why should I audit for it? Because its proxies are not neutral. Unpaid internships, relocation ability, prestige schools, and access to professional networks correlate with race and sex, which means a class filter can produce protected-class adverse impact that the four-fifths rule will catch even though class never appears in the data. The invisibility is the risk: because nobody is watching for socioeconomic filters, they tend to survive requirement reviews that catch more obvious problems.

Can I ask a candidate about accommodations during screening? No. The ADA bars medical and disability inquiries before a conditional offer, so questions about a candidate's health or accommodation needs during screening are legal exposure in themselves. What you can do is what Naledi does: offer every candidate a standard accommodation option at scheduling without asking why anyone takes it, and audit each requirement against whether it is an essential function of the job rather than a habit of how the role has been done.

We use a vendor's AI screener. Is the bias risk theirs? No. The EEOC has stated that an employer remains liable for adverse impact produced by a vendor's AI tool, so "the algorithm did it" is not a defense. If the tool is an automated employment decision tool used for hiring or promotion of New York City candidates, NYC Local Law 144 additionally requires an independent bias audit, publication of the results, and notice to candidates that the tool is in use. Run the four-fifths rule on the tool's outputs by protected group before deployment and on a schedule after.

Our headline hiring numbers look balanced. Are we fine? Possibly not, and this is the single-dimension anti-pattern. Naledi's overall gender hire rate looked balanced while women of color advanced at a markedly lower rate, which only appeared once she broke gender down by race. Compute selection rates for intersecting groups rather than headline demographics, at every funnel stage rather than at hire alone.