loading ...
new
now with built-in ats.
sign up freeback to blog
hiring
doesaihiringcauseracialbias,andhowdoesapertureavoidit
a 2025 stanford hai study tracked 3.4 million applicants across 1,700 job postings and found that ai screening tools discriminated against Black and Asian applicants, hidden inside averages that looked fine. here is exactly how that happens, made hands-on, and how aperture is built to avoid every mechanism behind it.
26%
of Black applicants applied to roles where the tool discriminated against their group
15%
of Asian applicants applied to roles where the tool discriminated against their group
tl;dr
stanford researchers found real ai screening tools recommending candidates at unequal rates by race, a disparity invisible until you look role by role instead of at the company-wide average. the cause is not a rogue algorithm. it is resume-shaped signals (names, schools, gaps, formatting) standing in for attributes the system was never supposed to use. aperture’s evaluation never sees those signals in the first place, and every score traces back to specific evidence from a structured conversation, so a disparity has somewhere to be found instead of nowhere to hide.
what the study actually found
before any numbers, here is the whole idea as one story. scroll through it slowly: 1,000+ applicants enter the pipeline, and the same pool gets sorted twice, by two very different machines.
how a resume screener sees them
how aperture sees them
identity signals used: 3
pass rate gap: 2.3x
identity signals used: 0
everyone interviewed
every tier mixed
name
ethnicity
zip code
3 of 8 advance
5 of 8 advance
7 of 8 advance
nameethnicityzip code
reasoning
communication
adaptability
ranked 1-8
ranked 9-16
ranked 17-24
one job posting, 1,000+ applicants in the pipeline.
they all land in the ATS. one pool, all mixed together.
a resume screener sorts them by name, ethnicity, and zip code. look at the groups it made: segregated, and not treated the same.
aperture deletes those signals before anyone is judged. it cannot sort people by what it never sees. everyone gets the same interview.
aperture ranks them again, this time only on the skills the job actually needs. look at every tier now: mixed.
that gap you just watched happen is real. in 2025, researchers at Stanford’s Institute for Human-Centered AI examined how algorithmic screening tools rated real job applicants across 150 employers and 11 industry sectors. the full study tracked 3.4 million people who submitted 4 million applications to 1,700 postings.
0%
of Black applicants applied to roles where the tool discriminated against their group
0%
of Asian applicants applied to roles where the tool discriminated against their group
the researchers applied the EEOC’s four-fifths rule, the same legal standard used in Title VII employment discrimination cases, which flags a hiring practice when one group is recommended at less than 80% the rate of the most-favored group. by that standard, they estimate roughly 40,000 additional applications would have advanced if every group had been recommended at the top group’s rate.
none of this required a company to intend discrimination, or even notice it. that is the part worth sitting with.
the pooling trick
the study’s most counterintuitive finding is also the easiest one to miss in a headline. when the researchers pooled recommendation rates across every job category a company posted, the numbers looked close to fair. the disparity only appeared when they broke the same data out by individual position.
a company screening for six different roles can clear the four-fifths bar on average while failing it badly for two of those roles specifically. an aggregate audit would never catch this. try it below with the same logic the study used.
try it yourself
same screening tool, same applicants. two ways of counting.
recommendation rate, all roles blended together81%
averaged across every job category, the disadvantaged group clears the EEOC’s four-fifths line. by this read, the tool looks fine.
illustrative of the pattern Stanford HAI’s 2025 study found across 3.4 million applicants and 1,700 postings: no adverse impact when recommendations are pooled across all positions, but the EEOC’s four-fifths rule fails for specific roles once you look position by position. dotted line marks the four-fifths threshold.
systemic rejection
the study surfaces a second effect worth naming: what the researchers call systemic rejection. when many employers use the same screening vendor, a single algorithmic pattern of disadvantage compounds across every company that candidate applies to.
the researchers found that 10% of applicants who submitted four applications through the same vendor were rejected everywhere, a rate higher than independent, uncorrelated hiring decisions would ever produce. one model’s blind spot does not stay contained to one company. it becomes a pattern that follows a candidate across their entire job search.
how this happens without anyone intending it
no serious vendor trains a model to see race. the mechanism is almost always indirect. resumes carry proxy signals, features that never mention a protected attribute directly but correlate with it closely enough to reproduce the same disparity. a name can carry inferred ethnicity. a school or a zip code can carry inferred socioeconomic and racial patterns baked in well before a candidate ever applies. an employment gap can carry inferred caregiving history.
a model trained to predict who gets hired, using historical hiring data shaped by decades of exactly this kind of human bias, will happily learn these proxies as useful signal, because in the training data, they were. the model is not malfunctioning. it is doing precisely what it was optimized to do, on data that was never neutral to begin with.
how aperture is built differently
aperture does not screen resumes before a candidate gets a real interview. every applicant who applies gets the same structured, adaptive interview, on their schedule, with no resume-based pre-filter deciding who is worth talking to. that single design choice removes the entire category of proxy signals the stanford study traces the disparity back to.
λ-CORE scores what a candidate actually said across six behavioral dimensions, cognitive reasoning, domain knowledge, communication, behavioral indicators, collaboration, and adaptability, and every score carries a confidence interval and traces back to specific evidence from the conversation. there is no single opaque number standing in for "good candidate." there is a rubric, applied the same way to everyone, with a paper trail.
the study’s core lesson is that aggregate fairness metrics hide per-position, per-group disparities. a system built for auditability has to make that breakdown possible on purpose, not bury it in a pooled dashboard number.
what this looks like in practice
no resume, name, school, or employment-gap signal ever enters the λ-CORE evaluation
every candidate answers the same structured, adaptive interview, on the same rubric
every score traces back to specific evidence from what the candidate said, not a single opaque number
outcomes can be broken out by role and by group, not just pooled, because that is exactly where this study found the disparity was hiding
talk to sales
see what your funnel looks like when identity signals are off the table. we will walk you through it on your own roles.
talk to uspublished july 8, 2026 by aperture team 10 min read
keep reading
related from aperture.
hiring
why hiring is broken
the structural failure nobody is fixing
read
hiring
why resume screening fails high growth startups
the case for moving beyond keyword filters
read
candidates
can ai interviews discriminate against you
a careful answer about fairness, what gets evaluated, and what gets ignored
read
from the product