How AI Screens Your Resume

Start with the number you have probably seen: that 75% of resumes are rejected by applicant tracking software before a human ever reads them.

There is no primary source for it. It circulates constantly and traces back to nothing that measures it.

What is documented is different and more useful. In a survey of about 2,250 executives across the US, UK and Germany, more than 90% of employers said they use software to filter or rank candidates, 94% of those hiring for middle-skills roles and 92% for high-skills roles, and 88% said they believed qualified high-skills candidates were being screened out because they did not match the exact criteria in the job description. That is employers describing their own process, and it says the criteria are the problem rather than the software being rogue.

This article explains what these systems actually do, which of them deserve the word AI, what they demonstrably cannot infer, what the law entitles you to, and what you can legitimately do. It does not contain tricks for beating a filter, for reasons covered at the end.

Three different things get called AI screening

They work differently and they fail differently.

Parsing and keyword search. An applicant tracking system reads your uploaded file and extracts it into database fields. One major vendor's documentation lists them: name, address, city, postcode, phone, email, education level, graduation date, institution, program, employment start and end dates, employer, job function, responsibility. Recruiters then search those records with ordinary boolean queries, using AND, OR, NOT and wildcards.

None of this is AI, and the distinction is not pedantic. European Commission guidance states that the defining characteristic of an AI system is "the capability to infer," and explicitly excludes systems that "follow predefined, explicit instructions or operations" and "operate based on fixed human-programmed rules."

Statistical ranking. A different layer compares your resume to a job description by mathematical similarity rather than exact words. The same legacy product that offers boolean search also offers what it calls conceptual search, which "broadens the search result by including documents that are similar based on concepts and proximities," so retrieved documents "do not need to contain everything entered as search criteria, only related concepts." The research literature studies this using embedding models, which compare documents as points in a numerical space.

Generative AI. A third and newer category, where a language model is asked to read resumes and judge or compare them. Whether an employer can tell that a resume was itself written by one is a separate question, covered in Lathic's article on detecting AI-written text.

These three behave so differently that findings about one do not transfer to another. They also sit under different legal definitions: New York City's rules cover "any computational process, derived from machine learning, statistical modeling, data analytics, or artificial intelligence," which sweeps in the boolean filter, while the European Commission's own guidance would likely place a pure rule-based filter outside the meaning of "AI system" altogether.

How you actually disappear

Nobody presses reject. That is the part the folklore gets wrong.

The mechanism is in the vendor documentation. Search criteria can be marked required, excluded or desired. Results are ranked by computed relevance, and there is a display cap. The documentation suggests showing "the top 300 by relevancy."

If you rank 400th, you were not rejected. You were never displayed.

There is one documented case of genuine automatic rejection, and it is worth being precise about. In the survey above, 48% of employers said that when hiring for middle-skills roles they screen out any resume with an employment gap of more than six months, automatically, on that basis alone. The figure is specific to that tier of hiring. The employer set the rule. The software executed it.

What it gets wrong

Four measured failures, each pointing at a different layer.

Layout defeats parsing. A 2025 preprint on resume extraction reports that roughly 20% of resumes use non-linear, multi-column layouts that disrupt reading order for standard extraction pipelines. Testing a purpose-built system on 13,100 real resumes drawn from one large employer's recruiting system, it reached an overall F1 score of 0.964 while managing only 0.846 on long free-text fields such as job descriptions, against 0.984 on dates. That corpus is mixed Chinese and English, so the exact figures do not transfer, but the pattern does: the prose part of your resume is the part machines read least reliably.

Ranking is unstable. In a peer-reviewed audit of three retrieval models across nine occupations, using over 500 real resumes, shortening resumes to job titles only increased the share of tests showing significant racial disparity from 93.7% to 96.2%. With less to go on, the name did more work.

The signal is partly an artifact. The same study found that equalizing how often names appeared in the training corpus reversed the direction of the effect. Black-associated names were then preferred in 51.9% of tests against 22.2% for White-associated names. Whatever the model was responding to, it was partly a property of the data rather than of the candidate.

Order matters when it should not. In work on generative models comparing pairs of CVs, several models preferred whichever candidate was listed first, with parity restored only when the order was counterbalanced.

What the bias research actually shows

Two findings that point in opposite directions, which is itself the finding.

A peer-reviewed 2024 audit of embedding-based retrieval models found they "significantly favor White-associated names in 85.1% of cases and female-associated names in only 11.1% of cases," with Black male candidates disadvantaged in up to 100% of cases in some tests.

A separate study of 22 generative models comparing CVs pairwise across 70 professions found the reverse on gender: all of them consistently favored female-named candidates, an effect that grew when an explicit gender field was added and became negligible when CVs were rated in isolation rather than in pairs. That second study is a preprint.

Different model families, opposite directions on one attribute. "AI screening is biased in direction X" is not a safe sentence. What is safe is that these systems respond to things that should be irrelevant, in ways that depend on the model and the framing.

Disability is the clearest case. In a peer-reviewed 2024 study, a CV enhanced with disability-related credentials such as an award, a scholarship and a panel membership was ranked above an otherwise identical control in only 15 of 70 trials. By condition: autism zero of ten, Deaf one of ten, depression two of ten, cerebral palsy two of ten, blindness five of ten. A version of the model configured with disability-justice principles raised this to 37 of 70, which is a significant improvement and also evidence that these behaviors are a property of configuration rather than a fixed fact about the technology.

Lathic's article on AI bias covers the general mechanism.

What the law gives you, by place

This changes fast and varies enormously. Five specifics, each dated.

New York City requires an annual independent bias audit of automated employment decision tools, with a publicly posted summary, and notice to candidates at least ten business days before use, including the qualifications being assessed and a stated right to request an alternative selection process. Enforcement began 5 July 2023. Two things temper it: the law "does not require any specific actions based on the results of a bias audit," and, as of September 2026, the enforcement agency had not published any penalty under it.

Illinois amended its Human Rights Act effective 1 January 2026 to prohibit AI use that results in discrimination in hiring and promotion, and to require employer notice when AI is used in a covered decision. Separately, since 2020, Illinois has required that before an AI-analyzed video interview an employer must explain how the AI works and what general types of characteristics it uses, and obtain consent.

California regulations effective 1 October 2025 treat automated-decision systems as capable of violating state fair employment law, require four years of retention of automated-decision data, and note that assessments eliciting disability information may be unlawful medical inquiries.

Colorado replaced its 2024 AI Act before it ever applied. The successor, signed 14 May 2026, requires disclosure of automated decision-making and, within 30 days of an adverse decision, a plain-language description of the decision and the technology's role. It applies to decisions from 1 January 2027.

The European Union classifies recruitment and candidate-evaluation systems as high risk, and gives an affected person the right "to obtain from the deployer clear and meaningful explanations of the role of the AI system in the decision-making procedure." That right runs against the employer, not the vendor. The high-risk obligations do not yet apply: a 2026 amendment moved them to 2 December 2027. The Act's general applicability began 2 August 2026, which is a different thing.

Federally in the US, the picture shifted in 2025. The EEOC's two AI technical assistance documents, the 2022 disability guidance and the 2023 adverse-impact guidance, are no longer on its website; both URLs returned a 404 when checked in September 2026. A government audit describes the wider pattern without naming a reason for any one removal, saying that several federal agencies that had issued guidance to employers "have either rescinded these prior efforts or are reassessing their alignment with the current administration's priorities."

Removal is not repeal, and this is the sentence to hold onto. The EEOC's worker-facing AI guidance of April 2024 is still published and states that federal employment discrimination laws "protect you when AI systems are used to discriminate against you," and that those laws "apply to the use of AI and other new technologies in employment just as they apply to other employment practices." Its enforcement plan for 2024 to 2028 remains in effect and still names AI screening as a priority.

Enforcement that has actually happened

One characterization point on the first. The EEOC's own description is of software programmed to reject applicants by date of birth, which is a deterministic rule rather than a learned model. It is the clearest documented case of automated screening harm, and it was not AI.

The Workday ruling is arguably more consequential. The court rejected the argument that the vendor was an employment agency but accepted agency liability, finding that a vendor carrying out delegated screening can be answerable, and that nothing supports treating software screening differently from human screening.

What you can legitimately do

Everything here is about being read accurately. None of it is about deceiving anything.

MIT's careers service publishes specific guidance. Use a .doc, .docx or .pdf file. Keep fonts at 10 point or larger, and use common ones such as Arial, Calibri, Cambria, Georgia, Helvetica or Times New Roman. Avoid graphics, icons and images. Avoid putting information in tables or text boxes. On reading order, it notes that text appearing in the wrong order "typically indicates that the setup of the document is incorrectly ordered," often because of text boxes or columns.

UT Austin adds: avoid tables, columns, graphics, objects, LaTeX and design-template sites, and use terms from the job description that match your actual experience.

There is a mechanical reason this works rather than a superstitious one. The vendor documentation states that "resume parsing has no impact on the formatting of a text (bold, italics, bullets)." Visual styling is discarded. What survives is the extracted text and its mapping into fields. A two-column layout is not penalized for being ugly; it is misread.

What you can ask for

Depending on where you are, more than most applicants realize.

If you are applying to a New York City employer, you are entitled to notice at least ten business days before an automated tool is used, to be told what qualifications it assesses, and to request an alternative selection process. The bias audit summary is required to be published, so you can read it.

If disability, religion, pregnancy or a related condition affects your ability to take an automated assessment, the EEOC's live guidance states that existing law may require the employer to provide a reasonable accommodation. Ask before the assessment, not after.

If you are in the EU, once the high-risk obligations apply, you will have a right to a meaningful explanation of the system's role in a decision that significantly affects you, and that right runs against the employer.

And everywhere: ask a person. Most of what matters in a hiring process still runs on human attention, and an applicant tracking system is a filing cabinet with a search box.

Why this article has no tricks in it

A large share of the demand behind this question is for keyword stuffing, white text and other ways to make a resume register as something it is not.

MIT's guidance closes that door directly, and it is worth quoting because it is the practical objection rather than the moral one: "Avoid spamming the ATS with keywords. You will still need to be able to account for everything indicated on your resume. Falsifying employment documents, including application materials, is both a legal and ethical breach and may result in dismissal from the search."

The mechanism is simple. A keyword that survives parsing puts you in front of a person who then asks about it. The distance between the resume and the interview is where the trick fails.