AI Resume Screening: How It Works, Where It Fails, and the Law

AI resume screening is the use of a model to read applications and rank, score or filter candidates before a recruiter looks at them. It is now the most common use of AI in recruiting: in iCIMS and Aptitude Research's April 2026 survey of more than 400 US talent acquisition leaders, 69% used AI somewhere in hiring and screening was the top use at 58%. In the same survey 45% had no governance framework for it. That gap, between adoption and control, is what this page is about.

The pages that rank for this topic are written by the vendors that sell the tools or by resume services selling the opposite promise, that you can beat the tool. Neither cites the studies that measured bias in the current generation of language models, the lawsuit that has turned a vendor into a defendant, or the state laws that took effect in 2025 and 2026. This guide does, with the date and sample on every figure, and it is written for the HR team deciding whether and how to use one.

What AI resume screening is and is not

Three different things get called AI resume screening, and the law treats them differently.

• Keyword and knockout filters inside an applicant tracking system. A rule such as "must hold a bachelor's degree" or "must have the words project management" removes applications before anyone reads them. This is automation, not AI, but it is the source of the "75% of resumes are rejected by a robot" claim, which traces to a 2012 vendor estimate and has never been measured. Harvard Business School's 2021 Hidden Workers report found 88% of executives saying their automated tools screen out qualified candidates, which is the better-sourced version of the same concern.
• Matching and ranking models that score each resume against a job description and order the list. These have been in applicant tracking systems for a decade and are what the regulations mean by an automated employment decision tool when a recruiter relies on the order.
• Large language model screening, where a general model reads the resume and the job and produces a judgment, a summary or a score. This is the newest and the one the 2024 bias studies tested.

The legal line is whether the tool "substantially assists or replaces" the decision in New York City's words, or "materially influences" it in Colorado's. A tool that summarizes a resume for a recruiter who then reads the resume is outside both. A tool whose ranking determines who gets a phone screen is inside both.

How many employers use it

Two readings matter. Use has roughly tripled in four years. And the volume problem that drives it is real: LinkedIn's data show applications per US role doubling since 2022, and 74% of talent leaders say candidates now use AI to write those applications. The model is screening resumes that another model wrote. Our HR technology statistics page tracks the adoption series.

Where it fails: the evidence

Name-based bias in current language models

The University of Washington study published October 31, 2024 is the most direct test of the newest tools. Researchers ran three large language models over more than 550 real resumes and 500 real job listings, swapping in 120 first names associated with white, Black, male and female people, for more than three million comparisons. The models preferred the resume with a white-associated name 85% of the time and the Black-associated name 9% of the time; male-associated names 52% of the time and female-associated names 11%. Black male names were never preferred over white male names. The resumes were identical apart from the name. This is the failure that a keyword filter cannot have and a language model does.

Learning the past

Reuters reported on October 10, 2018 that Amazon had scrapped an experimental recruiting model after finding it penalized resumes containing the word "women's", as in a women's chess club, and downgraded graduates of two all-women's colleges. The model had been trained on ten years of resumes submitted to a company whose technical hires were mostly men, and it learned that pattern as merit. Every ranking model trained on an employer's own hiring history carries the same risk, which is why the questions to ask a vendor below start with training data.

Qualified people filtered out

SHRM's 2022 survey found that among organizations using automated filters, 19% knew they had excluded qualified candidates by accident. Harvard Business School and Accenture's Hidden Workers report, September 2021, surveyed more than 2,250 executives in the US, UK and Germany and found 88% agreeing that qualified, highly skilled candidates were vetted out because they did not match the exact criteria. The report attributed the problem to over-specified filters, such as requiring a degree for a role that does not need one, and to gaps in work history being treated as disqualifying.

Opaque scoring

Bloomberg's 2024 test of a general language model ranking identical resumes with different names found similar name effects to the Washington study. Vendors' own bias audits published under New York City's law show impact ratios, but a 2024 study of 391 New York City employers by 155 student investigators found only about 5% had published the required audit, so there is very little public data on how deployed tools perform.

The law that applies in 2026

Federal discrimination law applies to a model's decision exactly as it applies to a recruiter's. Title VII's disparate impact provision makes an employer liable for a neutral practice that disproportionately excludes a protected group unless the practice is job-related and consistent with business necessity. The Uniform Guidelines on Employee Selection Procedures, 29 CFR 1607.4(D), treat a selection rate for any group below four-fifths of the highest group's rate as evidence of adverse impact. The Equal Employment Opportunity Commission's May 2023 technical assistance on this point was removed from its website in January 2025; the regulation was not. The Age Discrimination in Employment Act and the Americans with Disabilities Act apply too, and the ADA adds a duty to offer accommodations in the screening process, for example to a candidate whose disability produces a gap in work history.

State and local rules add notice, audit and record duties. New York City's Local Law 144 requires an independent bias audit within a year before use, a public summary and 10 business days' notice to candidates, enforced since July 5, 2023. Illinois HB 3773, in force since January 1, 2026, makes AI use with a discriminatory effect in recruitment and hiring a civil rights violation and requires notice. California's Civil Rights Council regulations, in force since October 1, 2025, make the screening vendor the employer's agent and require four years of records. Colorado's re-enacted law requires notice, a 30-day explanation after rejection and human review on request for decisions from January 1, 2027. The full table with citations is on our AI hiring laws by state page.

Mobley v. Workday

The case that every HR vendor is watching is Mobley v. Workday, Inc., No. 23-cv-00770 in the Northern District of California, before Judge Rita Lin. Derek Mobley, a Black man over 40 with anxiety and depression, alleged he was rejected from more than 100 jobs at employers using Workday's screening tools. The court let the case proceed against Workday as an agent of the employers, and on May 16, 2025 conditionally certified a nationwide collective of applicants aged 40 and over under the Age Discrimination in Employment Act. On June 22, 2026 the court granted in part and denied in part Workday's motion to dismiss the third amended complaint, letting California Fair Employment and Housing Act claims proceed on the basis that Workday's algorithmic tools materially participated in the decisions from its California headquarters, allowing an amended disability claim, and striking a theory that Workday was itself the employer. The case has not been decided on the merits. Its significance is procedural: a vendor whose tool screens applicants can be sued as the employer's agent, and a collective of every older applicant screened by the tool can be certified.

What candidates think, and what they do

Pew Research Center's survey of 11,004 US adults, fielded December 12 to 18, 2022 and published April 20, 2023, found 66% would not want to apply for a job with an employer that used AI to help make hiring decisions, 71% opposed AI making a final hiring decision, and 47% thought AI would treat all applicants more consistently than humans while 15% thought it would do worse. Women were more opposed than men, 70% to 61%. Those attitudes have not stopped behavior: LinkedIn's 2026 data show 81% of job seekers using or planning to use AI in their search, and talent leaders report AI-written applications as the norm. The practical consequence is that keyword matching is now matching the model's keywords, and the signal in a resume has fallen. Several employers have responded by moving weight from the resume to structured work samples, which is also the approach with the strongest validity evidence.

Questions to ask a vendor

1. What was the model trained on? If the answer is our hiring outcomes, ask how they tested for the Amazon problem. If the answer is a general language model, ask for name-swap test results like the Washington study.
2. Can we see the bias audit for our configuration, not the product in general? New York City requires an audit of the tool as used; a generic audit of the vendor's platform does not meet it if your weighting differs.
3. What are the impact ratios by sex and by race and ethnicity, and by age band? The four-fifths rule is the threshold the regulator and the plaintiffs' bar will apply.
4. What does the tool do with a gap in employment, a non-traditional credential or a disability disclosure? The ADA accommodation duty attaches to screening.
5. Can a recruiter see why a candidate was ranked where they were? Colorado's 30-day explanation and Connecticut's adverse-decision explanation both require you to say what the tool did.
6. Will you supply the developer documentation Colorado requires from January 1, 2027? Intended use, training data categories, known limitations, update notices.
7. What are the retention terms, and can we keep four years of records as California requires?
8. Who is liable if the tool discriminates? Under California's regulations the vendor is your agent; under Colorado's law an indemnity for one's own discrimination is void; Mobley shows the vendor can be sued directly. Get the contract reviewed with those three facts in mind.

How to deploy one defensibly

1. Use the tool to surface and summarize, not to reject. A tool that produces a reading order for a human who reads every application is outside most of the rules above and keeps the recruiter accountable.
2. If the tool does reject or rank for action, run the four-fifths test on your own applicant data every quarter, by sex, race and ethnicity, and age band, and keep the results.
3. Write the candidate notice once and use it everywhere: which tool, what it does, how to request accommodation, how to request human review.
4. Designate a reviewer with authority to override and a documented process for using it. Colorado defines this; it is good evidence under every other law.
5. Remove knockout criteria that are not job-related. The Hidden Workers report's core finding was that over-specification, not AI, does most of the filtering out.
6. Record the model version, the configuration and the date of each change. Four years in California, three in Colorado, one year of audit currency in New York City.
7. Put the whole arrangement in your AI use policy under the employment decisions section, with the named approvers.

Key takeaways

• 69% of US talent acquisition leaders use AI in hiring and screening is the top use at 58%, but 45% have no governance framework (iCIMS and Aptitude Research, April 2026).
• Three large language models preferred resumes with white-associated names 85% of the time and Black-associated names 9% of the time in a 2024 University of Washington test of three million comparisons on otherwise identical resumes.
• Amazon scrapped a recruiting model in 2018 that penalized the word 'women's'; 88% of executives told Harvard Business School in 2021 that their filters screen out qualified candidates; 19% of SHRM respondents in 2022 knew they had done so.
• Title VII disparate impact, the ADA, the ADEA and the four-fifths rule in 29 CFR 1607.4(D) apply to AI screening. The EEOC's 2023 guidance was removed in January 2025; the law was not.
• NYC requires a bias audit and notice; Illinois bans discriminatory effect and requires notice; California makes the vendor the employer's agent with four years of records; Colorado requires notice and human review from January 1, 2027.
• Mobley v. Workday has a nationwide ADEA collective certified since May 16, 2025 and FEHA claims proceeding since June 22, 2026 against the vendor as the employers' agent.
• 66% of US adults would not apply to an employer using AI in hiring decisions (Pew, December 2022), yet 81% of job seekers use or plan to use AI in their search (LinkedIn, 2026).
• The defensible deployment surfaces and summarizes rather than rejects, tests impact ratios quarterly, gives one notice everywhere and names a reviewer with override authority.