Ask on a Separate Voluntary Form the Decision-Makers Never See

Can you legally ask a candidate's race and gender?

Yes, under US federal law, with conditions attached to purpose and separation. The EEOC's stated position is that pre-employment requests for information are assumed to form the basis for hiring decisions, so an employer should not request information that discloses or tends to disclose race without a legitimate business need, and tracking applicant flow is named as such a need ¹. The 1978 federal selection guidelines expect those records to exist in the first place ².

The agency also names the mechanism. Where an employer legitimately needs the information, it "may obtain the necessary information and simultaneously guard against discriminatory selection by using a mechanism, such as 'tear-off' sheets," which separates the race-related information from the information used to decide whether a person is qualified ¹. The paper metaphor has outlived the paper, and it still describes exactly what a well-built form does. The same guidance adds a rail worth writing into your process: asking for race-related information on the telephone "could probably never be justified" ¹. So it does not belong in a recruiter screen, a scheduling call, or a chatty intro round.

On the recordkeeping side, the Uniform Guidelines on Employee Selection Procedures say each user should maintain and have available records disclosing the impact of its tests and other selection procedures on employment opportunities by identifiable race, sex or ethnic group, and should adopt safeguards to ensure those records are used for appropriate purposes such as determining adverse impact and are not used improperly ². Two duties in one section: collect it, and fence it.

State law adds requirements this article does not cover, and the design of the analysis itself is a conversation to have with counsel before it runs rather than after it produces something. Race and sex are the unusual case, asked for a purpose the agency itself names as legitimate. Which questions are actually illegal, and what replaces them applies the same qualification test to everything else the form asks. What follows is the mechanics, which are the part nobody writes down.

Where does the answer have to sit relative to the decision?

Outside every record a decision-maker reads. That is the whole constraint, and the other design choices follow from it: a separate form or step, storage keyed so the answer can be joined for aggregate analysis and never surfaced on a candidate profile, access limited to whoever runs that analysis, and no integration that syncs the value back into the hiring system as a field somebody can sort on.

What to check, in the order it usually breaks:

• The screen. Is the demographic block on the same page as the work history, or on its own step after submission?
• The candidate record. Does the answer appear anywhere on the profile a recruiter or hiring manager opens?
• Exports and reports. Does the answer travel in the CSV somebody pulls to build a shortlist?
• Access. Who can query the table, and is that list shorter than the list of people who make decisions?
• Retention. Is the answer kept as long as the analysis needs and no longer than the policy says?

Monday's work is one task: open your own application as a candidate, on a phone, and find out where the fields sit. Same-screen placement is the common default rather than a mistake somebody made, because it is what most application builders ship. Finding it is not a scandal, and fixing it is a configuration change plus a conversation about who has access.

Disability sits under a different rule and does not travel with race and sex. The Americans with Disabilities Act, enacted in 1990, bars an employer from asking a job applicant whether they have a disability, or about its nature or severity, before a conditional offer, while allowing questions about the ability to perform job-related functions ⁴. Voluntary self-identification for federal contractors runs under a separate regime with its own form. The practical consequence for an assessment is that the accommodation route has to work without anyone naming a diagnosis, which is how to offer accommodations without forcing disclosure.

Say what the data is for, and count the declines

One sentence of purpose, and decline-to-answer as a visible option you count as a category. An ornate explanation reads as a pretext and depresses the response rate the whole measurement depends on. A blank and a decline are different facts about a candidate, and a form that cannot tell them apart has thrown away the only evidence you had about who chose not to answer.

Wording that works, roughly: this is voluntary, it is not shown to anyone involved in hiring decisions, it is used only to check whether stages of the process are treating groups differently, and declining has no effect on the application. Wording that does not work: a paragraph about values, a link to a policy PDF, or a sentence promising the data helps the company build a better workforce. The first tells a candidate what happens to their answer. The second asks them to trust a stranger.

The cost of getting this wrong shows up downstream in other people's published work. Researchers collected every bias audit they could find published under New York City's Local Law 144 between July 2023 and early November 2024, and across the 116 audits in that sample, 83% reported missing race and/or sex information for some of the applicants the tool evaluated, several of them for far more applicants than they had data on ⁵. That gap is largely applicants declining to self-identify, which is precisely the point: the design of the invitation decides whether the measurement can exist at all. What to do with the missing rows is a separate question with a real answer, in reading an impact ratio next to its response rate.

Which report needs which population?

The annual headcount report counts employees; an impact analysis counts applicants and what each selection procedure did to them, and one table cannot produce both. Under 29 CFR 1602.7, every employer subject to Title VII with 100 or more employees files Standard Form 100, the EEO-1 ³, and the EEOC runs it as a mandatory annual collection reaching certain federal contractors at 50 or more employees as well, on a filing window the agency announces each cycle ⁶.

The categories overlap, which is the source of the confusion. The selection guidelines say their race, sex and ethnic classifications are consistent with the EEO-1 series ², so the same field labels serve both. Everything else differs: the population, when it is collected, how long it is held, who may read it, and what a difference in it would mean. A headcount report says who works here. An impact analysis says what your process did to the people who applied, which is the only one of the two that can tell you a gate is broken.

One more trap sits in the middle. A report of workforce composition is often produced to answer a fairness question, and it cannot: composition reflects years of hiring, attrition, promotion and the applicant pool itself, and a stage-level difference can be invisible inside it. The guidelines set an enforcement default worth knowing before the analysis is scoped: where the total selection process shows no adverse impact, federal enforcement agencies in usual circumstances will not expect a user to evaluate the individual components, and where the total does show impact, the components should be evaluated ². If the question is whether a particular screen is cutting one group harder, the answer lives in a per-gate selection rate with the right denominator, and when a vendor holds the data, in what to demand for an adverse impact audit of a screening tool.

So the deliverable from this article is not a form. It is a map of three things: where the answer is captured, where it is stored, and which report each copy feeds. Write that down once and the compliance question and the measurement question stop being the same conversation.

Read the evidence >