AI Resume Screening: How It Works, Where It Goes Wrong and How to Check It
Automated resume screening ranges from keyword filters to AI that reads each CV against the role. How each type works, the documented ways it filters out good people or repeats bias, and a short audit you can run before you trust any screening tool.
Ece Badak • Oct 18, 2026 • 7 min read
AI resume screening reads each application and compares it with what the role needs, so a recruiter starts with a ranked or labeled list instead of a pile of CVs. Good tools explain every result and leave the decision to a person. Weak tools filter on words and credentials, drop qualified people before a recruiter ever sees them, and can repeat the bias in the data they learned from. You can check which kind you have in an afternoon, with the audit at the end of this article.
How automated resume screening works
Most screening tools do three things: they turn each CV into structured data, compare that data with the role, and give the recruiter a result. They differ in how they compare.
| Type | How it compares | Typical result | Main risk |
|---|---|---|---|
| Knockout questions and filters | Fixed rules: a degree, a license, a minimum number of years, no gap over six months | Pass or reject | Rejects people who could do the job but miss one rule |
| Keyword matching | Counts words from the job post in the CV | A match score | Rewards CVs written to the job post; misses experience described in other words |
| Trained ranking model | Learns which past applicants were hired and ranks new ones by similarity | A score or star rating | Repeats whatever pattern the past hires had, including bias |
| AI that reads against criteria | Reads each role on the CV and checks it against written criteria | A label or score per criterion, with a reason | Only as good as the criteria; needs a person to review unclear cases |
Resume parsing comes first in all four. A parser pulls names, dates, job titles and education out of the file. Parsing alone does not screen anyone; it only makes the CV readable for the next step.
AI in recruiting is now common. In SHRM's 2025 Talent Trends survey of 2,040 US HR professionals, 43% said their organization uses AI for HR tasks, and recruiting was the area where organizations used it most (51%) (SHRM, 2025 Talent Trends).
Where resume screening goes wrong
Rigid filters drop qualified people
A large study by Harvard Business School and Accenture asked employers in the US, UK and Germany about their screening systems. 88% agreed that qualified high-skill candidates were vetted out because they did not match the exact criteria in the job description. For middle-skill roles, the figure was 94%. 48% of employers filtered middle-skill candidates for a gap of more than six months in their work history (Fuller et al., Hidden Workers: Untapped Talent, 2021).
A gap filter cannot tell a career break for a sick parent from a lack of skill. Neither can a degree filter tell a self-taught engineer from someone with no experience. Each rule looks reasonable alone, but together they hide many of the people a recruiter would want to call.
Models learn the bias in their data
In 2018 Reuters reported that Amazon had stopped using an experimental tool that rated applicants from one to five stars. The model was trained on ten years of CVs, most of them from men, and it learned to mark down CVs that contained the word "women's" (Reuters, 2018).
Newer language models have the same kind of problem. Researchers at the University of Washington changed only the names on more than 550 real CVs and asked three text models to rank them for over 500 job descriptions. Across more than three million comparisons, the models favored white-associated names 85% of the time and female-associated names 11% of the time. They never favored names associated with Black men over names associated with white men (Wilson and Caliskan, AIES 2024; UW News).
AI-rewritten CVs pass keyword filters
Many applicants now use AI to rewrite their CV for each job post. The rewritten CV repeats the job post's words, so a keyword filter scores it high. A recruiter then finds out on the phone that the experience behind the words is thin.
How to make AI resume screening more accurate
Accuracy here means two things: the people the tool ranks high can really do the job, and the people it ranks low really cannot. These habits help with both.
- Write criteria from the job's real duties. Ask the hiring manager what the person will do in the first six months. "Has run weekly stock counts" is a criterion; "detail-oriented" is not.
- Separate must-haves from nice-to-haves. Keep the must-haves short. Every hard rule you add is another way to drop a good candidate.
- Count experience, not job titles. A "Team Lead" title can mean many things. Ask the tool to judge what the person did in the role.
- Ask for a reason behind every result. A score with no reason cannot be checked. A reason for each criterion lets a recruiter see a wrong call in seconds.
- Send unclear CVs to a person. When a CV does not say whether someone meets a criterion, the tool should label it "needs review" and leave the call to a person.
- Compare a sample with your own judgment. Each month, have a recruiter screen 20 to 30 CVs without looking at the tool's result, then compare. Look closely at the cases where you disagree.
- Keep protected characteristics out. Do not write criteria about age, gender, family status, religion, ethnicity or health. If an old job description mentions any of them, remove it before a tool drafts criteria from the text.
Is AI resume screening legal?
In most places, yes, with conditions. Two rules matter most for employers in the US and Europe. This is a short overview, not legal advice.
- New York City Local Law 144. Since 5 July 2023, employers and agencies in New York City may use an automated employment decision tool only if it has had a bias audit within the past year. They must also publish a summary of the audit and notify candidates in advance (NYC Department of Consumer and Worker Protection).
- EU AI Act. AI that filters applications or evaluates candidates is a high-risk use. The full obligations for these systems apply from 2 December 2027. Under the GDPR, candidates already have the right not to be subject to a decision based solely on automated processing when it has a significant effect on them. The guide to AI interviews and the EU AI Act covers the dates and the questions to ask a vendor.
Both rules point the same way: know what the tool does, be able to explain each result, and keep a person responsible for the decision.
A short audit before you trust a screening tool
Run this with one open role and the CVs you already have.
| Check | How to do it | A good sign |
|---|---|---|
| Reasons | Open 10 results and read why each candidate got their label | Every criterion has a reason you can check against the CV |
| Rewritten CVs | Paste the job post's key phrases into a weak CV and screen it again | The result does not change much |
| Unclear cases | Screen a CV that leaves out one must-have | The tool says it cannot tell, instead of guessing |
| Names | Screen the same CV twice with two different names | The results are the same |
| Rejections | Find out what happens to a CV that misses a must-have | A person sees it before anyone is rejected |
| Your judgment | Compare 20 to 30 results with a recruiter's own screen | You agree on most, and the tool's reasons explain the rest |
If a tool fails the names check or rejects people with no human review, do not use it for live roles.
How Firstview screens CVs
Firstview's CV Screening reads each CV against criteria for the role. You paste the job description or describe the role in a sentence, and Firstview drafts the requirements, nice-to-haves and dealbreakers. You can change any of them.
Each criterion gets one of four labels: Matched, Not matched, Needs review or Not applicable, with a one-line reason. Only experience whose duties match the role counts, so "stock" in a cashier role does not count as inventory management. A CV that misses a dealbreaker is marked, and nobody is rejected automatically. When you change a criterion, every CV in the stage is evaluated again.
The same criteria then carry into the AI interview, so each candidate is judged by one standard from the CV to the final decision. Firstview does not verify identity documents or check whether a certificate is genuine. Those checks stay with your team.
