Standardization is a fairness technology.

The project of reducing bias in hiring can aim at two different targets: the evaluator or the procedure. Almost all of the spending goes to the first — training sessions, awareness programs, appeals to objectivity. Almost all of the well-supported results come from the second. Bias needs discretion to operate: a criterion that can move after the candidate has been seen, an interview that can wander, an impression that can stand in for a measurement, a final judgment that can reweight the evidence to fit a preference. The interventions with the strongest records work by closing those openings, not by upgrading the person standing in front of them.

The split is easy to miss because the first theory is more flattering. It says the process is sound and the people need calibrating, which implies a fix that can be delivered in an afternoon and imposes no constraint on how anyone actually decides. The second theory says the opposite: the people are roughly as good as people get, and the process is leaking. Its fix is less comfortable, because it withdraws latitude that evaluators experience as professional judgment. But it has one large advantage. A procedure, once changed, stays changed. It does not get tired, revert under deadline pressure, or make an exception for a candidate who reminds someone of a younger self.

This article walks through the evidence for the procedural theory: experiments in which evaluators moved the goalposts and pre-commitment stopped them, a natural experiment in which a physical screen was associated with changes in who advanced, and a meta-analysis showing that the same information predicts better when combined by rule than by feel. The pattern that emerges is mechanical rather than moral, and it lands somewhere useful: the openings close with tools an organization would want anyway. That convergence, not any single study, is the case for structured hiring.

Preference cannot act without an opening

Start with an audit question rather than an accusation: at which specific moments in a hiring funnel could a preference, conscious or not, change the outcome? Four moments recur across the research, and they share a shape. Each is a point where the process defers to discretion, where what happens next depends on who is judging rather than on what was measured. Figure 1 draws them as doors, because that is how they behave: open by default, each with a documented mechanism, each closable by a specific procedural choice.

The first door is the criteria themselves, when they are settled late. Uhlmann and Cohen (2005) demonstrated the mechanism in a set of experiments published in Psychological Science under a title that compresses the finding: constructed criteria. Evaluators comparing finalists inflated the importance of whatever credentials the favored candidate happened to have. When the preferred finalist was strong on formal education, education became the mark of merit; when the preferred finalist's strength was practical experience, experience did. The evaluators were not applying a standard unevenly. They were assembling the standard around a conclusion, after seeing the candidates. The same experiments contained the repair: when evaluators committed to the criteria in advance, the redefinition of merit disappeared.

The second door is the unequal interview. When each candidate faces a different conversation (different questions, different follow-ups, different difficulty), the resulting evidence cannot be compared, and a comparison happens anyway, on impression. The research on interview structure catalogs the elements that close this channel, from job-derived questions asked identically of everyone to anchored scoring of the answers (Campion, Palmer, & Campion, 1997); our companion article on why unstructured interviews fail traces both the collapse in accuracy and the repair in detail. The point here is the bias mechanism, not only the noise: an interviewer with a favorite can throw softballs without ever deciding to, and an unstructured format will not stop it.

The third door is the impressionistic screen: a human reading applications and deciding, on overall impression, who deserves a hearing. Field experiments have measured what travels through this opening. In the experiment our article on résumé screening accuracy examines in detail, Bertrand and Mullainathan (2004) sent employers matched fictitious résumés and found that identical qualifications drew meaningfully different callback rates depending on the name at the top. Nothing on the documents justified the gap, so the gap came from the reading. A screening step built on measurement rather than impression, a standardized assessment scored the same way for every applicant, gives the name no step at which to act.

The fourth door is the last step: combining the evidence. However disciplined the inputs, a panel that folds scores, ratings, and records into a holistic overall judgment has reopened every earlier channel at the finish, because holistic synthesis permits reweighting. Whatever the preferred candidate did well can quietly count for more. Kuncel, Klieger, Connelly, and Ones (2013) meta-analyzed the alternative: combining the same information by rule rather than by holistic judgment improved predictive validity by more than half. The bias reading of that result matters as much as the accuracy reading. A combination rule fixed in advance is a weighting scheme that cannot be bent toward a conclusion, which is exactly the promise a free-form debrief cannot make.

These channels do not require ill intent to operate; ordinary drift is enough, and in the constructed-criteria experiments the evaluators gave every sign of believing their own definitions of merit. They also compound. A preference admitted at the screen decides who reaches the interview, an unequal interview manufactures supporting evidence, and free-form combination ratifies the package at the end, each step laundering the one before it into something that looks like a record.

The inventory also converts an abstraction into an audit. Bias as a property of minds is unobservable at the level of any single decision, which is one reason arguments about it go nowhere. Channels are inspectable. An organization can check, process by process, whether criteria are written before candidates are seen; whether every interviewer asks every candidate the same questions; whether anything other than a scored measure decides who gets a hearing; whether the final weighting is fixed in advance or negotiated in the room. Each check has a yes-or-no answer, and every no marks a specific place where reducing bias in hiring is an available design decision rather than an aspiration.

The orchestra screen, told straight

The procedural theory has a natural experiment. Over a period spanning decades, American symphony orchestras changed how auditions ran: increasingly, candidates performed behind a physical screen that concealed them from the committee. Goldin and Rouse (2000), in a paper in the American Economic Review whose title states the design principle (orchestrating impartiality), studied what the change was associated with. Their finding, stated qualitatively: the adoption of screens was associated with meaningful increases in women advancing from preliminary rounds and in the share of women among new hires.

What made the study travel far beyond labor economics is its shape, which Figure 2 diagrams. The screen did not ask the committee to believe anything, notice anything, or overcome anything. It subtracted an input. Judgment continued at full strength; it simply attached to the playing rather than the player. That is the pattern this article is about, executed in physical form: identify the channel through which the irrelevant attribute enters the decision, and close the channel rather than coaching the decider.

The study also carries a caution that has to travel with it. Later re-analyses have questioned the precision and strength of some of its estimates; the honest reading is that the direction is supported and the magnitudes are debated. That is why this article quotes no percentages from it, and why Figure 2 carries the same caution in its caption. The debate subtracts less than it might, for a structural reason: the case for standardization does not rest on this one result. The constructed-criteria experiments and the callback audits demonstrate the same mechanism under randomized control, and the mechanical-combination meta-analysis measures the cost of the open final channel. A natural experiment earns realism at the price of control, which is why re-analysis is possible at all; the randomized studies trade in the opposite direction. The laboratory supplies the mechanism; the field supplies the demonstration. The screen is the memorable illustration, not the foundation, and that is the right posture toward blind hiring research generally: read each study as a probe of a mechanism, not as a promised effect size.

Procedures do not have to remember to be fair

Set against this evidence, the most common intervention looks misaimed rather than wrong. Awareness training addresses something real: assumptions exist, and unexamined ones influence judgment. But the evidence that a one-off training changes evaluative behavior, meaning what evaluators subsequently do under time pressure with real candidates and real stakes, is weak and mixed. That is not a reason for contempt, and this article is not an argument against training. The question is one of sequence: procedure-level interventions carry the stronger evidence, and they are the ones organizations tend to reach for last, because they constrain the organization rather than merely informing it.

The deeper reason is an asymmetry of maintenance. A trained evaluator has to apply the training at the exact moment discretion is exercised: late in the day, behind schedule, across from a likable candidate. A procedure has no such moment; an interviewer who intends to be fair and is running on fumes still asks the same questions in a structured hiring process, because the questions are the process. Procedures are also checkable in a way minds are not. Whether the anchors were used and the rule was applied sits in the record, while an awareness program leaves an attendance sheet, and so, when the intervention is procedural, reducing bias in hiring stops being a claim about interior states and becomes a property of the process that an auditor, a board, or a candidate can inspect.

The anatomy of a standardized selection process

What does it take to close all four doors? Not a slogan and not a single product, but five properties of the process, each doing a specific job. They are not new inventions. Four of the five generalize what the interview-structure literature has prescribed for decades, extended from a single conversation to the whole funnel; the fifth, documentation, is what makes the other four provable. Together they are what structured hiring means when it is applied end to end. Figure 3 maps the properties to the channels they close; several close more than one.

Same instrument for every candidate. Every applicant for a role faces the same questions, tasks, or measures, so every decision downstream rests on a common evidence base. This closes the unequal-interview channel, and applied at the top of the funnel it closes the impressionistic screen as well: when the first cut is made on a standardized measure, the first cut cannot be made on a name. It also changes what a debrief can be about, since a shared evidence base turns disagreement into a dispute over the same facts rather than a contest between different conversations. What that looks like operationally, with assessment preceding the résumé read, is laid out in the platform overview.

Anchored scoring. Responses are rated against scale points tied to concrete described behaviors instead of evaluative adjectives, so "strong" means the same thing across raters and across candidates. This is the pre-commitment that the constructed-criteria experiments found protective: the definition of merit is fixed before any candidate is seen, and cannot be rebuilt around a favorite.

Independent ratings before discussion. Each rater records a judgment before hearing anyone else's. This keeps the loudest or most senior voice from setting an anchor the room converges on, and it preserves genuinely separate readings of the evidence for the combination step to use. In practice it means written scores, entered before the debrief opens, not recollections assembled afterward. A panel that talks first and rates second has one opinion with several signatures.

Rule-based combination. The weights that turn scores into a recommendation are set when the role is defined, not when the finalists are known. This closes the final channel, the one the mechanical-combination meta-analysis measured: the rule that blocks post-hoc reweighting is the rule that preserved predictive validity in Kuncel and colleagues' data.

Documented basis. Every recommendation is traceable to what produced it: which instrument, which anchors, which rule. Documentation converts a hiring decision from an event into a record that can be reviewed, challenged, and audited. It also does quieter work with candidates, since consistency of administration is itself a rule of procedural justice (the perceived fairness of how decisions are made): applicants judge a selection system partly by whether everyone faced the same process (Gilliland, 1993). What candidates make of that judgment, and what it costs when they make it badly, is the subject of our companion article on the candidate experience of assessment.

Reducing bias in hiring is the same work as raising validity

Run down the five properties again and notice what else they are. Identical instruments, anchored scoring, independent ratings, and rule-based combination are the textbook prescriptions for making selection predictive, not just consistent. The flagship article in this series, on what actually predicts job performance, shows the pattern across a century of measurement: the methods at the top of the validity table are the disciplined ones, and the methods that lean on unaided discretion sit near the bottom. The interview is the cleanest single case, since adding structure to the same conversation is what turns one of the weaker predictors into one of the strongest.

The convergence is not a coincidence; it is one variable seen from two sides. Discretion is the opening through which preference enters, and discretion is also the opening through which noise enters: the mood of the rater, the order of the interviews, the confidence of the loudest voice. Close the opening and both leave together. Kuncel's result made the accuracy half measurable, and the constructed-criteria experiments made the preference half visible, but the intervention in both is identical. That is why an organization does not face a tradeoff between hiring fairly and hiring well. The evidence standards behind that claim are the ones documented on our science page.

The usual objection is that standardization deletes judgment. It relocates judgment. Deciding what a role requires, which behaviors anchor each scale point, and how the evidence should be weighted are acts of expert judgment, exercised where the research finds judgment most reliable: in advance, in the open, and applied to every candidate equally. Standardization is not the enemy of judgment; it is the delivery mechanism that makes judgment consistent enough to be fair. What it removes is the private, per-candidate variance that the four channels describe. A hiring manager's expertise shows up in the anchors and the weights, where it benefits every candidate, instead of in the room, where it benefits whoever the room happens to favor; the thinking is the same, done earlier, in the open, and once. Whether a process's outputs remain balanced is a separate, checkable question, and the standard monitoring test for it is the subject of our companion article on adverse impact and the four-fifths rule.

Change the procedure first: five commitments

The practical sequence follows directly from the mechanisms, and most of it is design work: decisions an organization can make this quarter, without waiting for anyone's judgment to improve. Each commitment names the channel it closes, because an intervention that cannot say what it closes is difficult to evaluate later.

• Commit to criteria before viewing candidates. Define what merit means for the role and write it down before the first application is opened. The constructed-criteria experiments are precise about the timing: pre-commitment is what eliminated the shift, and a criterion that exists only in memory is a criterion that can move.
• Standardize the screen before the interview. Most rejection decisions happen at the top of the funnel, where the evidence is weakest and the discretion is greatest. Replacing the impressionistic read with a standardized measure closes the channel where the callback experiments found the name acting.
• Score independently, combine by rule. Ratings entered before discussion, weights fixed before finalists are known. This is the cheapest of the five commitments and the one the meta-analytic evidence rewards most directly.
• Keep the record. Instrument, anchors, rule, and result, retained for every candidate. A reviewable decision disciplines the process that produced it and gives any later challenge something concrete to examine.
• Aim training at adoption, not substitution. Teach evaluators why the structure exists and how to run it well. Training works better as the manual for a redesigned procedure than as a replacement for one.

An organization that makes these five commitments has not purified anyone's instincts, and it does not need to. It has removed the points at which an instinct could become an outcome — which is the part of the problem an organization actually controls, and the only part it can verify.