AI Recruiting Academy › Module 3: Screening and interviewing with AI
The short answer
AI resume screening is software that reads applications and sorts them against the requirements of a job. Depending on the tool, it filters out people who fail set questions, scores each resume for fit, ranks the whole pile, or writes a short summary of why each person does or does not match.
It exists because application volume has exploded. It works best as a sorter that puts the strongest applications in front of you first. It is risky as a gatekeeper that rejects people nobody ever looks at, because screening models can carry bias and the law increasingly holds employers, and now vendors, responsible for the result.
Why are recruiters using AI to screen resumes?
Because there are too many applications to read. Ashby’s 2026 recruiter productivity report, based on more than 109 million applications, found the average recruiter now processes 291 applications per hire, compared with roughly 100 in early 2021. Only 3.6% to 4.7% of applications now lead to an interview.
Candidates are using AI too. LinkedIn told the New York Times it was receiving 11,000 applications a minute in 2025, up 45% in a year, and a Gartner candidate survey found 39% of applicants used AI during their application, most often to write the resume itself. When resumes are written by AI, they look more alike, and reading them by hand gets harder.
What are the types of AI resume screening?
“AI screening” covers four quite different things. Knowing which one a tool uses tells you most of what can go wrong.
| Type | How it works | Where it fails |
|---|---|---|
| Knockout questions | Rules you set, such as “Are you authorized to work in the US?” or “Do you hold a CDL?” | Not really AI. Fails when a rule is too strict or badly worded |
| Keyword parsing | Pulls titles, skills and dates out of the resume and matches words to the job | Misses people who describe the same skill differently |
| Matching models | Compares the meaning of the resume with the job, the same idea as semantic search | Can learn patterns from past hiring that you would never approve of |
| Language model scoring | A large language model reads the resume and writes a score and a reason | Can sound confident while being wrong, and can be swayed by names and wording |
Most modern applicant tracking systems now offer some mix of these. Some only summarize and rank. Others can move applicants to a rejected stage automatically. That difference matters more than any feature list.
Can AI resume screening be biased?
Yes, and the evidence is strong. University of Washington researchers tested three language models on more than 550 real resumes, changing only the name at the top. Across more than three million comparisons, the models favored white associated names 85% of the time and Black associated names 9% of the time, and they never favored Black male associated names over white male associated names.
Putting a person at the end does not fully fix it. A follow up UW study gave 528 people AI recommendations while they picked candidates. When the AI was moderately biased, people copied its bias. When it was severely biased, they still followed it around 90% of the time.
The lesson is not “never use AI”. It is that a human review only helps if the reviewer looks at the rejected pile, not just the AI’s shortlist.
What does the law say about AI screening?
This is an overview, not legal advice, and Module 4 covers the law in depth. Three things every recruiter should know today:
| Rule or case | What it means for you |
|---|---|
| US anti discrimination law | You are responsible for discriminatory outcomes even when a tool made the call. “The software did it” is not a defense |
| New York City Local Law 144 | An automated employment decision tool needs a bias audit within the last year, a public summary of the results, and notice to candidates 10 business days before use |
| Mobley v. Workday | In May 2025 a federal judge let an age discrimination claim against Workday’s screening tools proceed as a collective action. In 2026 the plaintiffs asked for class certification covering race, sex, age and disability. The case shows vendors can be sued, not just employers |
How should you use AI screening safely?
Rank, do not reject. Let the tool order the pile and summarize. Keep a person in front of every rejection, which is the checkpoint rule from Lesson 3 of Module 1.
Screen on the job, not on a past hire. Give the tool the real must haves for the role. Tools trained on “people like our best employee” tend to copy that person’s school, employers and background.
Read the reasons. A good tool shows which requirements each applicant meets. If the reason for a low score is vague or wrong, do not trust the score.
Spot check the bottom. Each week, read ten applications the tool ranked lowest. If you would have interviewed any of them, your setup needs work.
Keep knockout questions few and fair. Only use requirements the job truly needs, such as a license or the legal right to work.
Tell candidates. Gartner found only 26% of candidates trust AI to evaluate them fairly, while 52% believe it screens their application. A plain sentence on how you use AI earns more trust than silence.
What should you ask a vendor before you buy?
| Question | A good answer sounds like |
|---|---|
| Can the tool reject applicants on its own? | Only if you switch that on, and it is off by default |
| What does it look at when it scores? | Skills, experience and requirements from the job, with names and photos hidden |
| Has it had an independent bias audit? | Yes, here is the summary and the date |
| Does it explain each score? | Yes, every applicant shows matched and missing requirements |
| Is our candidate data used to train your models? | No, or only with your written agreement |
Quick answers
Does my ATS reject candidates automatically?
Many can, but only if rules or AI rejection are switched on. Check your knockout questions and any automatic stage moves in your settings. If you are not sure, ask the vendor to show you.
Can candidates beat an AI screener by stuffing keywords?
Old keyword tools could be gamed that way. Meaning based tools and language models are harder to fool with a list of words, but they can still be swayed by confident wording. That is one more reason to read the evidence, not just the score.
Can I paste resumes into ChatGPT to screen them?
Be careful. You would be sending personal data to an outside service, and a general chatbot has no bias testing for hiring. Module 4 covers what candidate data you can safely put into AI tools.
Is AI screening legal?
Yes, in most places, but with duties attached, such as bias audits and notices in New York City and growing rules elsewhere. The duty not to discriminate applies everywhere.