What AI Resume Screening Does to Your Pile Before You Open It

What Happens Between Submit and Shortlist?

Extraction, then comparison. The file is parsed into fields, which is where tables, multi-column layouts, graphics and unusual date formats lose information without telling anyone. The extracted text is then compared against the job description, the pile is ordered by similarity, and a model usually writes a short summary of each candidate for the recruiter. Nothing in that sequence reads a person. It reads a document about one.

The two layers have different ages and different failure modes. Parsing is a decade-old problem that no model release fixed, and its errors are silent: a candidate whose dates were rendered in a sidebar does not get an error message, and neither does the recruiter. Ranking is newer, noisier, and much easier to argue with, which is why it absorbs all the attention while the older layer does more of the quiet damage.

This stack now has a legal name. California's amended employment regulations define an automated-decision system to include screening resumes for particular terms or patterns, alongside analysing facial expression, word choice or voice in online interviews, and they took effect on 1 October 2025 ¹. Keyword screening and video analysis sit in the same regulatory category, whatever the difference in sophistication. When an ATS keyword filter stops separating anyone is the operational version of the same fact.

One cheap defense exists against the silent layer. Ask any vendor to show the parsed output beside the original file for a two-column resume, and keep an alternative route open for applications that fail to parse: a plain-text field, or an address a person reads. Parsing errors track document design rather than anything about the candidate, so they land hardest on people using an unfamiliar template, a translated CV or a phone.

Why Does the Ordering Reward Resemblance?

Because resemblance is the quantity it computes. The tool measures how closely a document matches the posting, and the posting is public, so the applicant pasted it into a model too. What used to be a weak signal about effort and fit is now a reading of which drafting tool the candidate used, applied evenly to everyone who used one.

That is not an argument that applicants are cheating. Tailoring an application to a posting is the behavior every careers page asks for, and a model doing the tailoring changes the cost rather than the intent. It is an argument that the resulting order carries less information than it did, and that the tool cannot tell you which part of the resemblance came from the candidate.

The authorship question cannot be recovered afterwards either. A well-known study estimated that between 6.5% and 16.9% of the text submitted as conference peer reviews could have been substantially modified by a model, and stated in the same breath that trends visible across a corpus may be too subtle to detect at the level of one document ². A pile can be characterized. A single application cannot be adjudicated. Whether an AI-written resume is a reason to reject anyone follows directly from that gap.

Use It to Organize, Never to Cut

Point the tool at the work with no judgment in it and keep the judgment. Summarizing each application against four questions written in advance, grouping the pile by which requirement is met, and extracting the two facts actually screened on are all safe uses. Turning the ordering into a threshold is not, because nobody in the room can say what the threshold sorted on.

The configuration that follows is short. Turn off any automatic rejection. Set the output to a summary and a grouping rather than a single order. Keep the requirement fields the tool extracts down to things a human wrote and can defend. Then pull thirty applications from the bottom of the ordering, strip the names and schools, and read them against the same written rule used at the top.

That last step is the whole audit and it takes an afternoon. You are reading for three things: applications the parser mangled, applications that met the requirement in words the posting did not use, and applications ordered low for reasons nobody can articulate once they are looking at them. All three are fixable. None are visible from the shortlist. Whether the resume screen is throwing away the wrong people is the question that audit answers.

What Do You Owe Candidates and Regulators?

Notice in some places, a posted audit in one, and an answer everywhere. Local Law 144 has barred employers, since 1 January 2023, from using an automated employment decision tool on a candidate for a job in New York City unless a bias audit was performed within the prior year, a summary of its results is posted publicly, and the candidate had at least ten business days of notice ³. That is a disclosure regime rather than a safety certificate.

Read a posted audit for what it contains, which is less than the name suggests. Under the city's own rules it computes selection rates and impact ratios by sex, race and ethnicity, and their intersections, and nothing else: not accuracy, not job-relatedness, not whether any individual was treated fairly. An independent auditor may exclude any category representing under 2% of the audit data, and an employer that has never used the tool may rely on an audit built on other employers' data, or on synthetic data where too little real data exists ⁴.

The vendor does not absorb the exposure either. In technical assistance issued in 2023 and withdrawn in January 2025, the EEOC answered whether an employer is responsible for a tool designed by someone else with "In many cases, yes", including where a vendor is acting as the employer's agent ⁵. That document is no longer current guidance, so read it as the agency's reading of ordinary agency principles rather than as law.

The concrete precedent is smaller than its reputation and still worth knowing. Three tutoring companies paid $365,000 to settle EEOC claims that their application software automatically rejected women aged 55 or older and men aged 60 or older ⁶. That was a hard-coded date-of-birth cutoff rather than a model, which is the point: the enforcement risk attaches to an automated rejection nobody could explain, whatever produced it. What to ask a screening vendor to verify a bias-testing claim is where that conversation starts.

See a sample report >