Most large employers, and a growing number of small ones, now use some form of AI resume screening. Software reads each application, compares it with the job’s requirements, and ranks or sorts candidates before a recruiter looks at them. Done well, it gives every applicant a consistent first read and saves recruiters hours of skimming. Done badly, it rejects good people for reasons nobody can explain. This post explains how AI resume screening works, how it compares with manual screening, where it goes wrong, whether candidates should opt out of it, and how recruiters can use it responsibly. We build AI hiring software ourselves, in RecruitSmartly, so we have seen both sides.
How AI resume screening works
Resume screening software has gone through two generations, and many employers still use a mix:
- Keyword filtering. Applicant tracking systems (ATS) have long been able to filter on keywords, job titles, years of experience and knockout questions such as “do you have the right to work here?”. It is fast and predictable, and it misses anyone who describes the same skill in different words.
- AI matching. Newer tools use language models to read the resume as a whole. They extract skills, experience and qualifications even when they are phrased differently, compare them with the role’s requirements, and produce a score, a ranking or a short summary of how the candidate matches.
The typical flow is: the resume is parsed into structured data, matched against the job’s criteria, scored or grouped, and handed to a recruiter, ideally with an explanation of why each candidate ranked where they did. Some systems then go further, scheduling interviews or running a first-round AI interview.
AI vs manual resume screening
| Manual screening | AI screening | |
|---|---|---|
| Speed at volume | Slow; recruiters often skim each resume in seconds | Every resume read in full, in seconds |
| Consistency | Varies by reviewer, time of day and pile size | The same criteria applied to every applicant |
| Context and judgment | Good at unusual careers and potential | Weaker; depends on how well the criteria are written |
| Source of bias | Individual reviewers’ preferences and shortcuts | Training data, criteria and proxies, applied at scale |
| Explainability | Often “not a fit” with no record | Can show which criteria matched, if the tool is built to |
| Cost per applicant | Rises with every application | Low once set up |
Neither is neutral. Manual screening is inconsistent and its biases are hard to see; AI screening is consistent, which means its mistakes and biases are consistent too, but it can also be measured and corrected. The best results come from combining them: AI for a thorough, consistent first read, and people for the decisions.
Where AI resume screening goes wrong
- Learning from biased history. A system trained to find people like past hires learns whatever shaped past hiring. Amazon dropped an experimental recruiting tool in 2018 after it was reported to penalize resumes that mentioned women’s organizations, because it had learned from a decade of mostly male hires.
- Proxies. Even without names or gender, signals such as employment gaps, particular universities, postcodes or clubs can stand in for age, disability, ethnicity or gender.
- Parsing errors. Two-column layouts, tables, text in images and unusual headings can scramble the extracted data, so the candidate is scored on a garbled version of their resume.
- Rigid criteria. Requirements copied from an old job description, or an exact number of years in one tool, screen out career changers and people who learned the skill another way.
- Gaming. Candidates paste keywords in white text, or hide instructions to the AI in the document. Good systems detect it; weak ones reward it.
- Automation bias. Recruiters who only ever see the top of the ranking never notice what the system is getting wrong at the bottom.
For candidates: should you opt out of AI resume screening?
Some employers let you ask for your application to be reviewed by a person instead of, or as well as, an automated system. In some places it is a legal requirement: New York City’s law on automated employment decision tools requires employers to notify candidates and say how to request an alternative process or accommodation where one is available, and in the EU and UK, data protection law restricts decisions based solely on automated processing that significantly affect people, and gives them the right to ask for a person to review such a decision. Elsewhere, it depends on the employer.
Whether opting out helps depends on your situation:
- It may help if your background is unusual: a career change, a long gap, experience gained outside formal jobs, or qualifications from another country that a system may not recognize. It also makes sense if you need an accommodation for a disability that affects how you apply.
- It may not help if your experience lines up clearly with the role. Many processes use AI only to sort applications, with people making every decision, and opting out can mean waiting longer for a reviewer with less time per resume.
If you are unsure, ask the employer: what the tool does, whether a person reviews every application or only the ones it ranks highly, and how to request human review or an accommodation. A good employer will answer.
How to write a resume both AI and people can read
- Use a simple, single-column layout with standard headings such as Experience, Education and Skills. Save it as a text-based PDF or Word file, not an image.
- Use the job’s own language where it is true. If the posting says “customer onboarding” and you did exactly that under another name, say both.
- Spell out acronyms once, and name tools and skills explicitly rather than implying them.
- Explain gaps and career changes in a line, so neither a system nor a person has to guess.
- Do not hide keywords or instructions in white text or metadata. Screening tools increasingly detect it, and recruiters treat it as dishonest.
For recruiters: using AI resume screening responsibly
- Write the criteria, not a profile. Define the skills and evidence the role needs, and screen against those, rather than asking the system to find people who resemble past hires.
- Let AI rank and summarize; let people decide. Avoid automatic rejection. If volume forces it for clear knockout criteria, such as right to work, keep it to those and review a sample of the rejected applications regularly.
- Insist on explanations. Each score should show which criteria were met and where in the resume the evidence is, so recruiters can check it and candidates can be given a reason.
- Test for adverse impact. Compare selection rates across groups where you lawfully have the data. In the US, a group selected at less than four-fifths the rate of the highest group is a common warning sign. Repeat the test whenever the criteria or the model change.
- Tell candidates, explain what the tool does, and offer a route to human review or an accommodation.
- Know the rules where you hire. Requirements differ: New York City requires an annual independent bias audit and published results for these tools, and the EU AI Act classifies AI used for screening job applicants as high-risk, with obligations on transparency, human oversight and record-keeping. This post is not legal advice; check with counsel for your locations.
- Handle the data carefully. Resumes are personal data. Set retention periods, limit who can see what, and check where your vendor processes and stores it, and whether it is used to train their models.
- Measure the outcome, not just the speed: how often recruiters agree with the ranking, how shortlisted candidates do at interview, and what candidates say about the process. Our guide to LLM evaluation metrics covers how to check AI judgments against people’s.
Choosing AI resume screening software
Questions worth asking any vendor, including us:
- What exactly does the system score, and can a recruiter see why each candidate got their score?
- Does it screen against criteria we define, or learn from our past hiring decisions?
- How has it been tested for bias, by whom, and can we see the results? Can we run our own tests?
- How does it handle unusual resume layouts, other languages and hidden text?
- Can it run alongside our applicant tracking system, or do we have to switch?
- Where is candidate data processed and stored, how long is it kept, and is it used to train models?
- What support does it give for candidate notices, alternative processes and audits where the law requires them?
Conclusion
AI resume screening is neither the end of fair hiring nor a neutral fix for it. Used to give every applicant a consistent, explained first read, with people making the decisions and bias measured regularly, it can be fairer and faster than a tired recruiter with a pile of five hundred resumes. Used to auto-reject anyone who does not resemble last year’s hires, it repeats old mistakes at scale. Candidates are entitled to ask how it is used, and in many places to ask for a person instead; employers should be ready with a good answer.
We built Hiring AI, based on RecruitSmartly, with these questions in mind: skill-based evaluation against what the role needs, with people making the decisions. It can be adopted as it is or adapted to your hiring process and applicant tracking system. For the wider picture of putting AI into business processes, see what AI workflow automation is.