AI Resume Screening: How It Works, Accuracy and Bias Risks

By Kevin Kim · Updated Oct 11, 2026

AI resume screening is software that reads each resume, converts it into structured data (titles, skills, dates, schools), and then scores or ranks applicants against a job's requirements so recruiters see the strongest matches first. It saves time on high-volume roles, but it only ever reflects the criteria and the history it was given. If either one is skewed, the ranking is skewed too, and at scale.

This page is for HR directors, talent acquisition leads and hiring managers who are deciding whether to switch screening on, or who already have and need to know what to check. It covers the mechanics, the documented failures, a step-by-step setup, and the safeguards. It sits inside the recruiting section of the AI for HR collection, so for the full hiring picture start with AI in recruiting. This is general information, not legal advice.

Key Facts

• Amazon's experimental recruiting engine rated candidates from one to five stars, was trained on resumes submitted over a 10-year period (most from men), and penalized resumes containing the word "women's"; the team was disbanded (Reuters, October 2018).
• In a University of Washington test of three language models comparing over 550 resumes with over 500 job listings, white-associated names were preferred 85% of the time versus 9% for Black-associated names, and Black male-associated names were never preferred over white male-associated names (University of Washington, October 2024).
• More than 90% of employers in a Harvard Business School and Accenture survey said they use their recruiting systems to filter or rank middle-skills and high-skills candidates, and 88% agreed that qualified high-skills candidates are screened out for not matching the job description's exact criteria (The Register, summarizing the 2021 report).
• Under the US federal "four-fifths" guideline, a selection rate below 80% of the highest group's rate is generally treated as evidence of adverse impact (29 CFR 1607.4).
• The European Commission classes CV-sorting software as high-risk AI, with those rules applying from 2 December 2027 (European Commission).

How AI Resume Screening Works

Most systems run in three stages, and a failure at any one of them reaches the final ranking.

Stage 1: Parsing. The software extracts text from PDFs, Word files or web forms and sorts it into fields: name, employers, job titles, dates, education, skills. This is the "Ingest" step in the ACE Framework, and it's more fragile than it sounds. Two-column layouts, text inside images, creative formatting and non-English characters can scramble fields, so a strong candidate arrives as a half-empty record.

Stage 2: Matching. The system compares the parsed record to the job. Older tools match keywords: does the resume contain "SQL" and "forecasting"? Newer tools use language models that read meaning, so "managed the regional budget" can count toward "financial planning" even without the phrase. This is the "Analyze" step.

Stage 3: Scoring and ranking. The system turns the match into a score, a grade, a shortlist or a pass/fail. Some use rules you write ("at least two years in a sales role"). Others learn patterns from past hires and past rejections. That second kind is the one that can absorb history you didn't mean to teach it. In ACE terms this is a document-review pattern feeding a scoring-and-routing pattern.

One distinction runs through everything that follows: a screen that applies criteria you can read and defend behaves very differently from one that infers its own criteria from past decisions. The first is auditable. The second needs testing, because nobody wrote down what it learned.

What the Documented Failures Show

The scrapped Amazon recruiting model

Reuters reported in 2018 that Amazon's machine-learning team had been building a resume-review tool since 2014 and found by 2015 that it wasn't rating candidates for software developer and other technical roles in a gender-neutral way. The models learned from 10 years of submitted resumes, most from men, and taught themselves that male candidates were preferable. They penalized resumes containing the word "women's," as in "women's chess club captain," and downgraded graduates of two all-women's colleges. Amazon edited the programs to be neutral to those terms, but people familiar with the effort said that gave no guarantee the system wouldn't find other discriminatory ways to sort candidates. The company disbanded the team (Reuters).

The lesson isn't that AI is uniquely biased. It's that a model trained to imitate past hiring will imitate past hiring, including its blind spots. Patching the one proxy you spotted doesn't remove the others. Amazon told Reuters the tool was never used by its recruiters to evaluate candidates, though it didn't dispute that recruiters looked at its recommendations.

Name bias in language-model ranking

University of Washington researchers tested three language models by ranking over 550 real resumes against over 500 job listings across nine occupations, varying only the first names (120 names associated with white and Black men and women). Across more than three million comparisons, white-associated names were preferred 85% of the time and Black-associated names 9%. Male-associated names were preferred 52% of the time versus 11% for female-associated names, and Black male-associated names were never preferred over white male-associated names (University of Washington).

Read the scope carefully. This was a controlled research test of general-purpose models, not an audit of any commercial screening product. What it shows is why you test your own configuration instead of assuming a vendor's model is neutral.

Quietly rejecting good candidates

Bias isn't the only accuracy problem. The Harvard Business School and Accenture "Hidden Workers" report found more than 90% of surveyed employers use their recruiting systems to filter or rank candidates, and 88% agreed that qualified high-skills candidates get vetted out for not matching the exact criteria in the job description. That figure was 94% for middle-skills workers (The Register). The report also points to rigid proxies, such as treating a degree as a stand-in for a trait or auto-rejecting an employment gap. That's a design problem you control, and it overlaps with skills-based hiring.

Real Examples by Function

The pattern: screening works best on verifiable, job-related requirements (license, certification, location, availability) and worst on soft or inferred judgments like "culture fit" or "potential."

How to Implement It Step by Step

1. Decide the screen's job before you buy. Write down whether it should only confirm minimum requirements, or also rank. Confirming is far easier to defend than ranking.
2. Write requirements as criteria, not preferences. Separate must-haves (a license the role legally requires) from nice-to-haves. Anything labeled a must-have becomes a knockout, so keep that list short. Writing job descriptions with AI covers how to keep those criteria clean.
3. Check your inputs. Parsing quality, consistent job titles and clean applicant records decide the output. HR data foundations for AI explains what to fix first.
4. Choose rules you can read where you can. Ask any vendor what the system scores, what it was trained on, and whether you can see why a given candidate ranked where they did. Our AI recruiting screener agent page shows what a well-scoped screening agent does and what it hands to people.
5. Test on a back-catalog. Run the screen on a past, already-decided requisition and compare its output with who was actually advanced and hired.
6. Measure outcomes by group. Compute the selection rate for each group at each stage and compare it with the highest group's rate. The four-fifths guideline is a screening signal, not a legal safe harbor, but a result below 80% is your cue to investigate (29 CFR 1607.4). Where you're in scope, run a formal bias audit as well (see the next section).
7. Keep a person in the decision. Have a recruiter review every rejection that rests on a knockout, plus a random sample of the "low" pile each week. The human-in-the-loop concept is the standard to design to.
8. Give candidates a way out. Tell applicants an automated tool is involved and offer an alternate route or an accommodation request.
9. Monitor and re-test. Job requirements, applicant pools and models change. Re-run the group comparison on a schedule, and every time you change the criteria.

The Rules That Apply

Regulation varies by location and is changing quickly, so confirm current requirements with counsel for each place you hire.

• New York City. Local Law 144 prohibits employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within one year before use, a summary of the audit is public, and candidates receive notice 10 business days before use. The Department of Consumer and Worker Protection began enforcing it on 5 July 2023 (NYC DCWP).
• European Union. AI systems used for employment, such as CV-sorting software, are classed as high-risk, with those rules applying from 2 December 2027 (European Commission).
• Other US states. Several states have passed or delayed their own AI-in-employment laws, and some have been revised after passage. Treat any date you read as something to verify the week you rely on it.
• Existing anti-discrimination law. Using software doesn't move the responsibility. Disparate impact principles apply to selection tools generally, which is why the four-fifths comparison above is the standard first check.

For a governance structure that works across jurisdictions, NIST's AI Risk Management Framework (released 26 January 2023, voluntary) organizes the work into Govern, Map, Measure and Manage (NIST). It's a useful skeleton for an internal screening policy.

Safeguards That Actually Work

Think of safeguards as layers, because no single one is enough.

• Narrow the scope. Screen on job-related, verifiable criteria. Drop fields that proxy for protected traits, such as names, photos, graduation years and addresses, before the model sees them.
• Prefer explainable output. If a recruiter can't see the reasons for a ranking, they can't catch an error. See explainable AI.
• Audit before launch and on a schedule. Group-level selection rates, at every stage, not only the final hire.
• Sample the rejections. Bias hides in the pile nobody reads.
• Protect the exit. An appeal path and an alternate application route.
• Own the decision. A named person approves every adverse outcome. Bias in AI explains the sources of skew if you need to brief your leadership team.

Common Pitfalls

• Treating "passed the vendor's audit" as "safe for us." An audit covers a tool on certain data. Your job families and applicant pool are different.
• Over-trusting the ranking. Recruiters read the top ten and ignore the rest, so the algorithm's order becomes the decision.
• Letting one proxy fix stand in for a bias review. Removing a single biased term, as Amazon did, doesn't stop other correlated signals from doing the same work.
• Training on your own history. If past hiring favored one profile, a model that learns from it will too.
• Knockout questions with no review. Automatic rejection on a work gap or a degree is where qualified people disappear.
• Skipping candidate notice. In some jurisdictions that's a violation on its own.
• Never re-testing. A screen that was fine at launch can drift as requirements change.