Daniel Chait, chief executive of Greenhouse, which builds applicant tracking systems, has a name for what AI has done to hiring: an "AI hopelessness loop." Both sides have a problem, each tries to solve it with AI, and the attempt makes the situation worse. Candidates believe machines are screening them, so they use machines to write their applications. Employers drowning in hundreds of near-identical submissions run those submissions through AI scoring to tell them apart. The sharpest evidence that the machinery is not working comes from a company that tested its own: at Doist, when closed roles were fed back through automated screening, the people the company had actually hired did not make the shortlist.
Start with what job seekers are optimizing against. Jodi Beggs, a data scientist looking for work, ran her materials through a system that compared them with a job description. It told her two pages of resume was too many. It penalized her for using a middle initial in one document and not in another. When she swapped the word "percent" for the % sign, her score went up. Beggs did not say which tool she used; the best-known service of this kind is Jobscan. Her reasoning was straightforward: if the job is to please the robots, then even someone who dislikes AI-written text should consider that machines might like it.
Jobscan and its competitors rest on one assumption — that HR departments use AI inside applicant tracking systems to score and rank applications automatically, that only the top 10 to 20% of applicants get looked at, and that the remaining 80% never get a chance. That assumption is what candidates are paying to defeat, with keyword stuffing and formatting changes.
The assumption is sometimes true. In some organizations automated sorting really does run; in others, humans handle the process from start to finish. At its simplest, an applicant tracking system just collects applications, helps HR staff read them, and tracks a candidate through interviews to hire. Fuller products add onboarding management and automated AI interviews. Chait says job seekers believe a great many myths about all this: two systems do not necessarily behave the same way, the built-in AI can work one way today and differently tomorrow, and whether AI is involved at all depends on the specific product and which features a team paid for and switched on.
None of which changes candidate behavior. As long as applicants believe AI is doing the screening, they will try to beat it with AI of their own.
There are good reasons not to hand the outcome of a job search to a tool that promises to optimize you for a system you cannot see. One recruiter demonstrated that with extremely low Jobscan scores he still collected 12 interview invitations and a job offer. And Jobscan costs $30 to $50 a month, which makes the incentive structure worth stating plainly: the company has no commercial reason to want a subscriber to stop looking for work quickly.
The surrounding market makes everything worse. Openings and hires are both scarce. Fraudulent and fake listings have become a visible enough problem that several states are considering new laws to deal with them. Candidates and employers alike describe an erosion of trust — applicants sink hours into applications and hear nothing back, and applying turns into a black box. Employers describe receiving hundreds of applications that read almost identically, which is precisely why they load them into the AI scoring built into their tracking platform: to find the differences faster. The more AI is used this way, Chait argues, the greater the demand for more AI, and it helps no one.
Reporting for this story involved dozens of recruiters, HR leaders and small business owners who do their own hiring. Some admitted to using automated candidate ranking. Others flatly denied it. The split tracked neither company size nor application volume. What seemed to decide it was corporate culture and attitude toward the technology.
Kim Jones, vice president of human resources at Toshiba, says every application there is read by a person. She has no objection to candidates using AI to improve a resume or cover letter, but does not think it will get anyone past an applicant tracking system; selection comes down to the role's requirements, salary expectations and a few other factors, including whether the person is eligible for rehire. Where she does notice unwanted AI use is in interviews: a pause, then the sound of typing, then an answer that runs far too long.
Doist, a small fully remote company that hires worldwide and therefore receives a flood of applications per opening, went further and ran the experiment. Nadia Vatalidis, who leads people operations there, took roles that had already been filled, uploaded the job descriptions and every saved candidate file, and asked whether the tracking system's scores would have picked the same people. She wanted to know who the AI would have shortlisted and whether that list matched the candidates the team actually interviewed. In the two cases tested, the eventual hires were not on the AI's shortlist. There was some overlap among interviewees, but the new employees who were performing well about six months in would not have survived automated screening.
Then there is what a determined candidate can build on the other side. James Jacobsen started looking for work five months ago and found that searching and applying consumed as much time as a full-time job. A design specialist, he knew AI would not match him on creative work, but he had experimented with it enough to know where it was useful. He used Claude and ChatGPT to edit his materials and check them against job descriptions. It changed nothing.
So he stopped working on documents and started working on the process. He had Claude organize and track the entire search end to end: scan listings, analyze job descriptions, log them. He built a detailed scoring system accounting for job type, seniority and salary requirements, with different thresholds for onsite, hybrid and remote roles, and the assistant surfaced the roles worth applying to first. When he passed on a listing he wrote down why, so that when the same role was reposted weeks later the system could remind him. He spent less time searching and found more suitable openings. Several employers showed interest. No offers. He then asked Claude and ChatGPT to critique his portfolio and spent six hours rebuilding it. Two days later a prospective employer called. Also no offer.
What Jacobsen built is the inverse of an applicant tracking system — a search management system for the applicant — and it is the most instructive failure in this story. It worked. The process improved, the targeting improved, the time cost fell. And the outcome did not move. That is the thing the optimization industry cannot admit: for a large share of candidates, effort is not the binding constraint, so tools that multiply effort return nothing. Chait says candidates like Jacobsen keep asking how to increase their effort further, and that doing more of the same is the wrong answer. He also notes that this is the first time he has seen both sides of the market unhappy at once, which is his own evidence that the system is not working.
Consider who is making that diagnosis. Chait runs a company that sells applicant tracking systems — the layer that employers are stacking AI scoring onto. His account of the loop is accurate and his advice to candidates is sound, and he is also describing a demand curve his own product category sits on. That does not make him wrong; it makes the loop structurally durable, because the diagnosis and the supply are coming from the same place.
The question nobody in this story can answer is the simplest one: which employers actually run automated ranking. Dozens of hiring professionals were asked and the answers split with no pattern by size or volume. Vendors will not publish it. Employers are not required to disclose it. Candidates therefore spend money and hours optimizing for a system whose presence in any given application they have no way to verify — and the optimization industry's entire pitch depends on that uncertainty never being resolved. A disclosure requirement would be a duller reform than most AI policy proposals and would do more to deflate this market than anything else on the table.
The advice from the people closest to the problem is, tellingly, all pre-AI advice. Chait suggests researching companies you would genuinely want to work for, including ones that are not famous. Jones says she has nearly stopped seeing cover letters, which means sending one with a resume is now a way to stand out. Networking — finding contacts and referrals — is the other underrated skill in this market. Chait's closing message to job seekers is that the problem is not necessarily them; it is a system that works badly.
So the moves that still work are the ones that do not scale, in a market where both sides just finished scaling everything else. Candidates are being told to go slower and more human at the exact moment the volume of applications makes slow, human review the first thing employers automate away.