An estimated three out of four resumes submitted to a job posting are filtered out by software before a human recruiter ever opens them. The exact figure moves depending on which study or vendor you ask, but most estimates converge above 75%, which means the entity making the first, most consequential cut in modern hiring is very often not a person at all. It's an applicant tracking system, running keyword matches and scoring rules against a resume that a human will, in a large share of cases, never see.
That alone would be worth scrutiny. What makes it more than a process story is where these systems get their rules from. Many are configured or trained using historical hiring data: the resumes of people who were previously screened in, interviewed, and hired for similar roles. If that historical pattern already reflects bias, whether that's a preference for candidates without employment gaps, a skew toward certain schools, or something closer to the outright discrimination described below, the software doesn't correct for it. It learns the pattern and applies it forward, at the scale of every resume that touches the system, rather than the scale of one recruiter's individual judgment on one afternoon.
A case that made the pattern explicit
In 2023, iTutorGroup agreed to pay $365,000 to settle a lawsuit brought by the U.S. Equal Employment Opportunity Commission. The allegation wasn't a vague claim of algorithmic unfairness; it was specific and mechanical. The company's hiring software had been programmed to automatically reject female applicants aged 55 and older, and male applicants aged 60 and older, filtering out more than 200 qualified candidates on that basis before a human ever reviewed their applications. There was no ambiguity to litigate over intent versus effect: the rejection criteria were coded directly into the system.
The iTutorGroup case is useful precisely because it's the clean version of a much messier problem. Most algorithmic bias in hiring software doesn't look like an explicit age cutoff written into the code. It looks like a scoring model that quietly downweights resumes with certain patterns, patterns that happen to correlate with age, gender, disability, or other protected characteristics, without anyone deciding to build it that way. iTutorGroup shows what happens when that same logic is stated outright instead of buried in a model's weights.
A test case still working through the courts
A more open-ended version of the same question is now moving through federal court in Mobley v. Workday. The plaintiff alleges that Workday's AI-powered screening tools, used by many of its corporate customers to filter applicants, disadvantaged candidates on the basis of race, age, and disability. A federal judge allowed key discrimination claims in the case to proceed, rejecting Workday's argument that it couldn't be held liable as merely a vendor rather than the employer making the actual hiring decision.
Central to the allegations is the idea of a proxy variable: a factor that isn't itself a protected characteristic but correlates closely enough with one that using it produces the same discriminatory effect. Employment gaps are the clearest example named in the case. A gap in someone's work history might reflect a layoff during a downturn, a disability, a period of caregiving, or a return to education, and several of those reasons skew toward older workers, women, and people with disabilities. A screening algorithm that treats a gap in employment as a negative signal doesn't need to know anyone's age or disability status to produce an outcome that disadvantages those same groups. That's the legal theory at the center of the case, and it's still being tested rather than settled.
Why the pattern outlasts any one lawsuit
Neither case proves that all ATS software discriminates, and neither is fully resolved: iTutorGroup settled without admitting the software's design reflected intentional discrimination in the legal sense, and Mobley v. Workday is still an active case rather than a finding of fact. What both cases demonstrate, regardless of their individual outcomes, is that the filtering layer sitting between 75% of resumes and the humans meant to evaluate them is not a neutral piece of infrastructure. It is a set of rules, written or learned, and those rules inherit whatever assumptions went into building them. A biased recruiter affects the applications they personally review. A biased filter affects every application that passes through it, silently, before anyone with judgment gets a look.