Decline to Self-Identify

The questionnaire comes at the end, which is the interesting part. You have uploaded the resume. You have retyped the resume into eleven boxes because the parser mangled it. You have named a salary, given three references and agreed to terms nobody reads. Only then, when you are tired and invested and would very much like this to be over, does the application ask what you are.

The page is titled “Voluntary Self-Identification of Disability.” In the corner, in small type, it carries a federal form number (OMB control number 1250-0005) and an expiration date. A federal document that asks you to declare your condition has, itself, a shelf life. There is a companion page for race and ethnicity and another for veteran status. Each offers three ways out of the question, and the third is the most philosophically loaded sentence in American employment law: “I do not wish to answer,” or, in the older and better phrasing, “decline to self-identify.”

Consider what that phrasing assumes. It presupposes a self, already identified, sitting somewhere behind the applicant and being withheld. It treats the answer as a fact you possess and are choosing not to hand over, like a phone number. This essay argues three things. In several important cases there is no such fact waiting to be disclosed. The form is not reading the answer off the applicant but writing it onto them. And this is happening tens of millions of times a year without anyone having decided to do it.

The best vocabulary for this problem was coined by Darwin, in a letter to Hooker in 1857, and he was complaining about colleagues. Those who make many species, he wrote, are the “splitters,” and those who make few are the “lumpers.” He meant it lightly. It has outlasted almost everything else in the letter, because the distinction it names turns out to be permanent.

Here is how a species is anchored. Somewhere there is a type specimen: a single dead individual in a drawer, usually in London or Washington or Leiden. The name is attached not to the population in the field but to that one carcass. Everything else is argument. Whether the birds on this island and the birds on that island are one species or two is settled by committees, which meet, deliberate and vote, and have been known to reverse themselves twice within a working lifetime. The world’s bird species number about eleven thousand, except that until the unified AviList of 2025 the two main checklists disagreed with each other by several hundred. A well-known 2016 analysis argued that under a stricter definition of species the true figure is closer to eighteen thousand. That is not a rounding error. It is seven thousand kinds of animal appearing and vanishing according to which definition is in fashion, while not one bird in the field moves an inch or notices anything at all.

That could be the whole essay, but it is not, because of one difference the bird cannot supply. The bird does not read the checklist. It has never been told of its promotion and has no view on the committee. It will breed or fail to breed on exactly the terms it would have used had the volume never been printed. The philosopher Ian Hacking called these indifferent kinds. Quarks and mosses and weather systems do not consult their classification, which is exactly why the classification can be treated as a description of them.

The other sort he called interactive kinds, and the entire difficulty lies in the difference. When you classify a person, the person may find out. Having found out, they may agree, resist, organize, apply for something, change their account of their own history, or simply begin to notice things about themselves that they had never put under a heading. The classified thing changes, which changes what is true of the class, which changes the classification, which can be read again. Hacking called this the looping effect, and he spent a career on the observation that in the human sciences the instrument is never merely an instrument. A drawer that talks back is not a drawer.

This brings us back to the last page of the application, and to the fact that it does not ask an abstract question.

The current disability form does not say “Do you have a disability?” and leave you to it. It prints a list that includes the following. Alcohol or other substance use disorder. Autoimmune disorder, including lupus, fibromyalgia and rheumatoid arthritis. Blind or low vision. Cancer, past or present. Celiac disease. Deaf or hard of hearing. Depression or anxiety. Diabetes. Epilepsy. Gastrointestinal disorders. Mental health conditions, including bipolar disorder, schizophrenia, post-traumatic stress disorder and obsessive compulsive disorder. Missing limbs. Nervous system conditions, including migraine headaches. Neurodivergence, including attention deficit hyperactivity disorder, autism spectrum disorder and dyslexia. Short stature. Traumatic brain injury.

Read that as an applicant rather than as a compliance officer. It is not the language of measurement. Measurement asks a question and waits. This list enumerates and invites recognition, which is how a screening questionnaire works, and how a clinician works at intake. The difference is that a clinician has fifteen minutes, a medical history and a duty of care. This list is administered by a payroll vendor to a person who has just typed their own name for the fourth time and would like to know about the health plan.

So does it do anything? Has anyone ever read that page and gone looking for a diagnosis? The honest answer is that the mechanism is well understood and the direction of travel is not in serious doubt. Migraine, depression and adult attention deficit are conditions with real biology, wide prevalence and a lot of diagnostic latitude at the margins. A person’s decision to seek assessment is strongly influenced by whether the category has been offered to them in a form that seems to fit. Hacking’s original case study was multiple personality, where the description’s availability and its circulation through clinics and paperback memoirs came before, and produced, the statistics. You do not have to accept that reading in full to accept the general point, which is modest: categories that are printed, official, repeated annually and attached to a benefit will be lived in more than categories that are not.

The same is true, more quietly, of the demographic page. The cleanest example in American statistics is the word “Hispanic,” which was manufactured. Before the mid-1970s the federal government had no such category, and Mexican Americans, Puerto Ricans and Cuban Americans did not generally think of themselves as one kind of people. A congressional act in 1976 and a directive from the Office of Management and Budget (OMB) in 1977 created the classification for statistical purposes. Activists, broadcasters and marketers then had an interest in filling it. Within two decades a real shared identity existed where there had been none, with its own politics and its own media market. Cristina Mora’s history of the process is called Making Hispanics, and the title is not a provocation. It is a description.

The process has not stopped. In March 2024 the OMB revised the standards again, made a single combined race and ethnicity question the default, and added a Middle Eastern or North African category. Several million people became a new kind of American by administrative revision, having done nothing whatever, like a bird promoted to full species while asleep. The difference, once again, is that they can read the notice.

There is a wrinkle in the practice that makes the argument easier to see. It is common enough that most people who have applied for anything in the last decade have met it without noticing. Some employers ask twice. Before the federal pages arrive, in the body of the application itself, among the custom questions the company has built in its own applicant tracking system, there is a second set of demographic questions, written in-house.

Point72, the investment firm, did this for years, and the evidence is public. Pull any of its Greenhouse job pages (Greenhouse is the hiring software) out of the Internet Archive, whether a snapshot from November 2021 or one from January 2025. Partway down the form, after school and degree and years of experience, there is a section headed “Diversity.” It explains that the firm invites the information to track the effectiveness of its diversity initiatives in connection with its recruitment outreach, that providing it is voluntary, and that the answers will not form part of the application. Then it asks. I identify as: male, female, non-binary, decline to self-identify. Are you Hispanic/Latinx? Ethnicity/Race: African American or Black, Asian, Multiracial, Native American Indian or Alaska Native, White or Caucasian, each of them tagged “not Hispanic or Latinx,” and at the bottom, “I choose not to provide this information.”

To the firm’s credit, the fields are optional, the disclaimer is prominent, there is no disability question anywhere in the block, and nothing about any of it is unlawful. I grant all of that without grudging it, because it is true and because the firm went to some trouble to make it true. But then: why ask?

The number already exists. The federal form, three pages later, on the same afternoon, under the same roof, collects the same information and produces the same count, and the firm is obliged to file it whether it wants to or not. So the second asking cannot be about whether the number exists. The stated purpose says what it is about, and it is worth reading slowly: the “effectiveness of diversity initiatives in connection with recruitment outreach.” That is attribution. It is not “how many” but which channel, which campus, which posting, which stage of the hiring funnel, at what cost per applicant. It is ordinary marketing analytics, competently specified, and the units are people.

Two things follow. First, none of it requires asking before the offer. Ask afterward and you learn the conversion rate you actually care about, from the population you actually hired, at no cost to the several thousand people who were never going to work there. What the pre-offer question buys, and the only thing it buys, is the top of the funnel: the denominator, meaning the ratio, meaning the slide.

Second, and it matters more: the federal separation between the answer and the hiring decision is built into the system. It is a legal requirement, it has an inspector behind it, and it survives a change of management. The in-house separation is a sentence. “Will not form part of your application” is a promise written by the party making it, about data sitting in that party’s own system, enforced by nobody, audited by nobody, and reversible by a configuration change that notifies no one. Both boxes look the same on the screen. Only one of them is solid, and the applicant has no way to tell which.

That is why this is a useful example, because now look at the two menus side by side. The federal form offers “Hispanic or Latino” and does not recognize non-binary. The in-house form offers “Latinx” and does. The federal form treats Hispanic origin as an ethnicity crossed with a separate race question, or, after the 2024 revision, as a single combined question, depending on which version of the form the vendor has loaded. The in-house form does something else again. In one sitting, on one afternoon, the applicant is asked what they are twice, by two taxonomies that disagree, and must produce an answer to each.

Two rulers that give different lengths mean that at least one of them is not a ruler. If these categories were descriptions of a prior fact about the person, waiting to be reported, this could not happen. The fact would not change between the eleventh and the fourteenth question of the same form. What changes is the menu, and the applicant obligingly changes with it. That is the entire argument of this essay arriving unbidden in the middle of a hedge fund’s application, needing no interpretation. It is also why the answer to “why ask?” matters less than the fact of the asking. Whatever the question does to the person on the other side of it, a second question, in a different vocabulary, ten minutes later, does it again.

Larger employers go one better and put the jurisdiction in the label. American Express, among others, runs an application whose demographic block has a field headed “Gender (International),” with its own menu beneath it, sitting alongside the American questions rather than replacing them. The administrative reading is dull and entirely sound: a company that hires in dozens of countries needs a field that works where the reporting rules are not American, so it builds a second gender question and marks its scope in the name. The unintended reading is better. A label that qualifies a category by territory has conceded that the category has a territory. That raises a question the form cannot answer and has not noticed it is asking. If a candidate is internationally male, may he identify differently locally? Is the local answer allowed to differ from the international one? And if it is, which of the two is the man? Nobody in the building will ever see either answer, so nothing turns on it, which is the only reason the contradiction is allowed to sit there in production.

What has actually crossed the border here is not a view about gender. It is a field, a menu and a reporting obligation, which is the usual cargo. (My essay “The Org Chart Travels” is about this: the org chart travels, and the argument stays home.) And once the qualifier is on the label, there is no principled place to stop. The form has no box for interplanetary genders and no reason within its own logic to refuse them. Should the org chart keep going, some vendor will ship a field headed “Gender (Off-World),” and a compliance team will approve it on a Thursday.

Not every employer is so careful, and this is where the practice stops being merely interesting. An in-house demographic field marked required is a different animal from an invitation.

For race, the legal position is untidy. Collecting it is not unlawful in itself, since employers must collect it in order to file reports at all. A mandatory field is best understood not as a crime but as a document your future lawyer will wish did not exist.

For disability, it is not untidy at all. The Americans with Disabilities Act bars disability-related inquiries before a conditional job offer, without qualification. The only reason a company may put the disability question to an applicant at the application stage is a narrow regulatory exception for federal contractors. The conditions are that the question be an invitation, that it be voluntary, and that it use the prescribed form rather than a rewritten one. Build your own version, in your own words, and require it, and you have not merely bent the exception. You have removed what made asking lawful in the first place and kept the asking.

And who complains? That is the honest end of it. The remedy exists, it is real, and it has teeth. But reaching it requires being the applicant who writes in to object, at the moment they are hoping to be liked, about a field on a form, to a firm that has not yet decided whether to interview them. Nobody is fooled about what that costs. To object is to volunteer a different and far more consequential piece of information about yourself than any of the boxes asked for: that you are the sort of person who reads the form. A right that can be exercised only by someone who has already given up what they wanted is not quite a right. It is a fire exit with a camera over it, and the building is not lying when it says the door is unlocked.

The natural next question is who reads the answers. The answer is nearly nobody, which is both the reassuring part and the funniest part.

The responses do not reach the hiring manager. They are not supposed to, they are generally walled off in the applicant tracking system, and this is the one part of the apparatus that works exactly as designed. They are aggregated. They become a count in an EEO-1 filing (the annual federal workforce report to the Equal Employment Opportunity Commission), or a VETS-4212 report (the equivalent for veterans), or an analysis measured against the federal contractor goal that seven percent of each job group be individuals with disabilities, or against the veteran hiring benchmark that the Labor Department revises each year and sets at around five percent. The count is filed. In the ordinary case it is then looked at by an in-house employment lawyer, once, in a bad mood, and by nobody else ever.

There is something genuinely comic in the scale of this. Tens of millions of times a year, at the exact moment a person is most anxious to be liked, an institution stops and asks a question about the makeup of their being. Then it takes elaborate technical measures to make sure no one in the building will ever see the answer. It has the structure of a sacrament conducted by a congregation that has mislaid the god but kept the vestments and the calendar. The asking has become the point, because the asking is what can be audited.

It is worth noting how differently the same problem has been handled elsewhere. Britain asks all of this and more, with a longer list of ethnic categories and better manners about it, usually on a monitoring form physically detached from the application so that the separation is visible rather than merely asserted. France takes the opposite view and very nearly forbids the exercise. Collecting data that reveals racial or ethnic origin has been restricted there since the data protection law of 1978, and when the legislature tried to open the door to ethnic statistics in 2007, the Constitutional Council shut it. The French position is that a republic of undifferentiated citizens cannot sort its citizens by kind without ceasing to be one, and that counting the thing is a way of sustaining it. Two liberal democracies, with the same history in view, reached opposite conclusions about whether the census is the cure or the disease. Both cannot be right. The interesting possibility is that both are describing real costs and simply pricing them differently.

And then, on the same page, sitting there unremarked, is the veteran question. It is the most useful thing on the form, though not for the reason it was put there.

Protected veteran status has an unglamorous document behind it. The DD-214 is a certificate of release or discharge from active duty, which records the dates, the character of the service and the campaign badges. There is an enlistment on one side of it and a separation on the other, both events in the world with witnesses. The categories are dull and legally precise: disabled veteran, recently separated veteran, active duty wartime or campaign badge veteran, Armed Forces service medal veteran. When the federal government asks you this, it is asking for information it already holds.

Nobody has ever read the phrase “protected veteran” on a job application and gone quiet for a moment, wondering. Nobody has seen that box and thought, I have always been tired in a way I could not name, and now that I read the list, perhaps. Nobody joins a veterans’ organization, adopts the vocabulary, revises their account of their twenties, and gradually becomes the kind of person the classification describes. The category does not loop, because it is not a kind of person at all. It is a record of an act, chosen, dated and stamped, and the person is only the party who performed it.

Put that box next to the other two and the asymmetry is plain in a way no amount of argument about the other two could establish alone. One question on that page has an outside referent that would survive the abolition of the question. The others, to varying degrees, do not. This is not a claim that race and disability are unreal, which would be both false and stupid. A severed spinal cord is not a social construction, and neither is the history of the American South. It is a narrower claim: the veteran box points at a filing cabinet, while the disability box points at a mirror, and the applicant is asked to look into it and report.

The objection to all of the above is decisive as far as it goes, and anyone making this argument who does not concede it in full is not arguing in good faith.

The counting worked. That is the concession, and it is not a small one. You cannot detect a pattern you have refused to measure, and the entire structure of American antidiscrimination law after 1964 depends on being able to compare a workforce with a population and notice a gap. Griggs v. Duke Power (1971) is not a hoax and was not a technicality. A company was using a diploma requirement and an aptitude test that screened out Black applicants at a wildly disproportionate rate, for jobs the requirement had nothing to do with. It was stopped, and it was stopped by arithmetic. Without denominators there is no disparate impact, and without disparate impact the only remaining remedy requires proving intent, meaning proving what was in a man’s head, which means almost never proving anything. The French position, whatever its philosophical elegance, has left France conspicuously unable to answer basic factual questions about its own society, and a state that cannot see a problem is not thereby a state without one.

Grant all of that, and something survives, because the concession is about counting and the objection is not. A count can be assembled from records, from payroll, from observation, from what an institution already knows about the people inside it. The Census Bureau has spent decades refining methods for exactly this, and does not generally need to hand each person a printed list of conditions and invite them to recognize themselves in it. What the last page of a job application does is different in kind from measurement. It is enumeration delivered to the subject, at the least appropriate moment, by an institution with power over them, with a menu attached.

Separate the two and most of the difficulty dissolves. Keep the audit, which is defensible and probably necessary. Lose the catechism. If the aggregate is all that matters, and the wall between the answer and the decision-maker is real and solid, then the burden of self-examination placed on the individual applicant is pure deadweight: a cost imposed on the one party to the transaction who gets nothing from it. The obvious partial fix, which costs nothing, is to move the whole business to after the offer, where the count is just as good and the applicant is no longer being asked to audition.

The committee that splits a species in two has not done anything to the birds. It has done something to the ornithologists, who will now go into the field carrying a distinction they did not have last season, and who will find it, because that is what a distinction is for. The literature that follows will report a great many sightings of a bird that was previously not there to be seen. None of this is fraud. It is what happens when you improve the resolution of an instrument: you get more objects, the objects are real, and they were not previously separate.

Do that to people and you get the same result with one addition, which is that the people are also holding the checklist. The form at the end of the application does not manufacture a disabled person out of nothing, and it does not create a woman or a veteran or anyone else. What it reliably produces is applicants who have looked. That is a smaller effect than the alarmists claim and a considerably larger one than the drafters intended, and it is entirely unmeasured, because the one thing this apparatus is not set up to count is what the counting does.

None of this is an argument for abolition, and a reader who wants one will have to find it elsewhere. It is an argument for knowing what you have built. The people who wrote these forms believed they were installing a thermometer. They installed something closer to a mirror with a caption, hung it in a corridor that every working adult in the country walks down several times a decade, and then, having taken great care that no one would look at the readings, declared the matter handled.

And at the bottom of the page, patiently, in the same typeface as the rest: “decline to self-identify.” It is offered as the modest option, the one for people who would rather not make a fuss. It is in fact the only line on the form that gets the metaphysics right, since it is the only one that allows for the possibility that the question has no answer waiting behind it. That it is presented as a refusal rather than as a position tells you everything about which way the institution thinks the description runs.