The sorting machine: recruitment’s race to the bottom

By Iain, 10 Jul 2026

The sorting machine was built to cope with a genuine problem. Where it went wrong was in mistaking efficiency for effectiveness, in assuming that the ability to process resumes fast meant the ability to evaluate candidates well. The fix is not a smarter machine but a more modest one, paired with humans who are given the time, the tools, and the structured information to do what they were always better at.

If you’ve been involved in any part of the recruitment process, the last 18 months have been dispiriting, to say the least. Hiring became a closed loop. Employers adopted applicant tracking systems (ATSs) to handle volume, and candidates responded by tailoring their resumes to get past these systems’ mechanistic filters.

Employers then tightened the filters, prompting even more AI-driven tailoring. And here we are in 2026, with AI-generated resumes being scored by AI-powered screening tools, both sides optimizing for keyword overlap, while the question of whether anyone is actually well-suited to the job recedes into the background. This piece explains how the machine works, what it does to the people on both sides of it, and what employers should do about it.

The numbers tell a clear story. Jobscan’s 2025 ATS Usage Report found that 97.8% of Fortune 500 companies use an ATS, and the global market hit $17.22 billion in 2025. The typical online job posting now attracts 250 or sometimes many more applicants, of whom roughly 2% will get a serious interview. A recruiter reviewing hundreds of resumes for one role cannot read them all with equal care. Something had to give. What gave, unfortunately, was nuance.

How the machine reads a resume

The process begins with parsing, which converts a resume into structured data. The system extracts fields such as job titles, company names, dates, and skills, then stores the result for comparison with the job description. Modern parsing systems achieve roughly 87% accuracy in data extraction, versus 96% for human reviewers. That gap means roughly one in ten data points is wrong or missing, and creative layouts, tables, multi-column formats, and non-standard section headings make it worse. Over 75% of organizations report struggling with parsing failures caused by the variety of resume formats.

After parsing comes keyword matching. Most deployed ATS platforms still rely on exact and fuzzy matching rather than semantic search, so if the posting says “Google Analytics 4” and your resume says “GA4,” it may not register. The result is a process that treats language as a proxy for competence, and it is a terrible proxy. Matched resumes are scored, ranked, and sent to human reviewers if they exceed a threshold. Those below it are parked or auto-rejected.

Then come the knockout filters, binary gates that eliminate candidates outright. No degree? Excluded. Fewer than five years of experience? Excluded. Resume gap longer than six months? Excluded. ATSs auto-reject 89% of candidates who fall even one year short of the stated experience requirement, with only 11% of systems allowing any flexibility. These are not AI decisions in any meaningful sense, but crude Boolean operations configured by humans and executed without discretion.

The blunt instrument and its consequences

The Harvard Business School and Accenture “Hidden Workers” study found that 88% of employers surveyed believed their ATS excluded qualified, high-skilled candidates because those candidates did not match the exact hiring criteria. As Joseph Fuller, co-author of the study, told the Harvard Gazette, these filters “aggregate in a way that causes a large number of people who might be 80%, 90% of the way home to being qualified, to fall out of the candidate pool, having never been assessed by a human being.” The study estimated that 27 million people in the US alone are “hidden workers” who are systematically filtered out of the workforce. These include caregivers, veterans, people with disabilities, the formerly incarcerated, and immigrants.

It is easy to cast the ATS as a callous automaton, but the ATS does not reject candidates autonomously. It organizes and filters based on criteria defined and managed by people. The software is a tool, and the problem is how the tool is configured. But that distinction is cold comfort when the criteria are set too tightly, the filters stack cumulatively, and the people configuring them rarely see the consequences on the other side.

Fuller’s team noted that most job descriptions are rarely updated with input from the people currently doing the work, and tend to accumulate requirements like Tupperware lids in a kitchen drawer. Employers modify old descriptions when posting new roles, tacking on additional skills and credentials rather than writing new specifications. Many of these are also drafted primarily using AI tools.

The result is a Frankensteinian document that is, at best, a hodgepodge of previous versions, or, alternatively, demands a unicorn when the organization needs a competent horse. How many hiring managers write a job description as a first-order task as opposed to an administrative annoyance to be tidied away as quickly as possible? This is a parlous state for arguably one of the most important classes of document in any organization.

The waste is asymmetric too, because some candidates take the wish list literally. Tara Mohr’s research into the famous statistic that women apply only when they meet 100% of the listed qualifications, compared with roughly 60% for men, found the explanation was not confidence but a reasonable belief that the stated requirements were real and that applying without them was a waste of time. Every preference written as a requirement deletes the most conscientious applicants first, before the parser gets anywhere near them. A lazy job description is not just a paperwork failure; it is an unconscious filter running upstream of the ATS filter.

From the applicant’s chair, the dominant experience of all this is silence. You apply for 80 roles, and what comes back is a standard rejection if you’re lucky, or, more often, nothing at all. (I would argue that if a company’s processes are so slipshod that they lack the wherewithal to send an automated rejection, then you’ve probably dodged a bullet.)

So candidates adapt to the reader they know they have. They flatten their resumes into a single column, strip the design, swap their own vocabulary for the job description’s, and delete the career detour that made them interesting because a gap may trip a filter somewhere. People are homogenizing and re-formatting their working lives for a machine’s convenience. The cruelty of the system is entirely procedural. Nobody decided to be explicitly cruel because nobody actively decided anything, which is precisely the problem.

The arms race nobody wins

According to SHRM’s 2025 reporting, an estimated 40% to 80% of job applicants now use AI to write resumes and cover letters. As Scott Holloway wrote in Crain’s Chicago Business, the average hiring manager is no longer reading applications so much as grading AI against AI. In one campus hiring drive, a recruiter reported receiving over 2,300 near-identical resumes. When every submission hits every keyword, the system’s ability to distinguish genuine expertise from competent prompting collapses.

The recruiter is the other casualty of this arms race, and the less discussed one. Most people go into recruitment because they like people, but the job has become exception-handling for a queue. The recruiter facing 2,300 near-identical resumes is not making 2,300 judgements. They are performing a ritual of diligence over output no human can meaningfully differentiate.

Trust erodes in both directions. Candidates assume no person will read what they send, so they let a model write it. Recruiters assume a model wrote whatever they receive, so they let a model read it. Both assumptions are increasingly true, and each tightens the other. The 2025 SHRM Benchmarking Survey found that average cost-per-hire and time-to-hire have both risen over the past three years, a period that coincides with increased use of generative AI in the hiring process. More technology, worse outcomes.

The law is coming

Unsurprisingly, the courts have noticed. In Mobley v. Workday, a plaintiff who applied to more than 100 jobs on Workday’s AI-powered screening platform and was rejected every time without interview has been allowed to proceed as a nationwide collective action. The court was unambiguous, writing that “drawing an artificial distinction between software decisionmakers and human decisionmakers would potentially gut anti-discrimination laws in the modern era.”

Other legislation is arriving from every direction. The EU classifies recruitment screening as high-risk under the AI Act, with full obligations from December 2027. California has imposed disparate-impact liability and four-year recordkeeping since October 2025. Illinois added effect-based discrimination liability and a private right of action from January 2026. The UK’s ICO published Recruitment Rewired in March 2026 and found most employers failing the test, with tools making the decisions and humans rubber-stamping them at a click. The jurisdictions may disagree on timelines and enforcement, but they agree on the substance. Tell people when a machine is judging them. Keep records of what it did, test it regularly for bias, and make human review real rather than ceremonial.

What to do now

The first job is an inventory, because the most damning thing the ICO found was not malice but ignorance. Employers genuinely did not know they were making automated decisions, since ranking and knockout features shipped on ATS platforms bought years ago for interview scheduling were switched on. Map every point in the funnel where an algorithm’s output influences a decision, and accept that under every current framework the deployer carries the obligations even when the vendor built the tool and swears compliance is their department.

The second is to make human review mean something. A reviewer who clears 200 AI rankings an hour is not reviewing anything. That configuration is just automation with a witness. Meaningful involvement requires someone competent, with genuine authority to overturn the machine, and enough time per candidate to exercise it. Building review processes where judgement is real is a retention measure for recruiters as much as a defense for lawyers.

The third is notice and records. Illinois requires notice now, the UK requires it before the decision, and the EU will require it at scale from December 2027. Burying it in a 20-page terms of service does not count. California’s four-year retention rule for ADS inputs, outputs, and selection criteria is the strictest current benchmark and a sensible default everywhere. Test before deployment, retest on a schedule, and document the results.

The fourth is the vendor relationship. The Mobley agency theory holds that exposure lies at both ends of the contract, so neither side can indemnify its way out. Get actual details on training data, bias testing methods, and parsing accuracy stipulated in the contract, rather than just asserted in the sales deck.

And the fifth, running through all of it, is to switch off scored auto-rejection. Knockout checks on genuine non-negotiables are fine. Automated rejection on a match score below a threshold will not survive scrutiny.

A better model

The natural instinct when faced with a broken system is to build a smarter version of the same thing. Improve the NLP. Add semantic matching. Train on better data. This is the pitch from every ATS vendor in 2026. Unfortunately, it completely misses the point. The problem is not that the technology is too crude, but that the ambition is wrong. Expecting an algorithm to make qualitative judgements about human potential from a two-page document is not optimistic. It is delusional.

A more honest framing accepts that AI cannot select the best candidates. It can, if designed with care and restraint, exclude the clearly unsuitable ones and leave the remaining pool for human judgement. This means rethinking the ATS as a light-touch triage filter rather than a ranking engine. Red-flag exclusion, covering verifiable, non-negotiable requirements such as a current medical license or right-to-work, is where automation works well.

It should not govern “preferred” qualifications or degree requirements copied from a 2019 job description. The critical design principle is no automated ranking within the surviving pool, because the moment you rank, you reintroduce every bias the system learned from historical hiring data, every keyword-matching incentive, and every reason for candidates to game the system. A ranked list, plus a tired, overstretched human, is how “decision support” becomes automated decision-making in practice.

With the clearly ineligible filtered out, give recruiters better tools for the remaining pool. Structured skills assessments (a 2022 SHRM survey found that 79% of HR professionals rated them as important as, or more important than, education or experience). Asynchronous video responses give access to the human behind the resume without the scheduling overhead. Work samples and structured application questions are harder to fake than keyword-optimized text. And AI-assisted summarization that prepares materials for the human who will make the decision makes the system act as a research assistant rather than a judge. A recruiter reviewing a triaged pool of 80 candidates, with summary cards, assessments, and video responses, is conducting recruitment. A recruiter clearing an AI-ranked queue is doing quality assurance on a machine, badly, at speed. Only one of these jobs is something employees can be proud of.

The sorting machine was built to cope with a genuine problem. Where it went wrong was in mistaking efficiency for effectiveness, in assuming that the ability to process resumes fast meant the ability to evaluate candidates well. The fix is not a smarter machine but a more modest one, paired with humans who are given the time, the tools, and the structured information to do what they were always better at.

Somewhere in the rejection pile of every ATS in the world, there is a candidate who would have been excellent at the job. But they used the wrong synonym. They had a six-month gap while caring for a parent. Their resume was formatted in two columns, and the parser mangled their work history into nonsense. The system processed them in under a second, and three floors up, a recruiter who would have liked them sat, clearing a ranked queue, never knowing they existed. Two people that the machine kept apart. That is not a technology problem but a design choice, and we can make a different one.