Why AI Is Changing How Companies Hire

You've almost certainly done this: found a posting that matched you well, tailored an application, heard nothing back — not a rejection, not an auto-reply beyond the first one, nothing — and eventually noticed the same listing was still live four months later. That experience has a name now, and a number attached to it. A 2024 study built on a novel dataset of Glassdoor interview reviews found that roughly one in five online job postings shows the statistical fingerprint of a listing nobody intends to fill — kept open to build a candidate pipeline, signal growth to investors, or quietly benchmark the market. The pattern concentrates in large firms and high-skill industries, the opposite of where you'd expect sloppy recruiting to cluster [1].

Here's the part that complicates the obvious villain story: the people posting those listings are drowning too. At one recruiting software firm's own customers, the average recruiter now juggles far more open requisitions than four years ago, reviews well over double the applications per year, and works on a team that's shrunk by roughly a quarter — while the number of people actually hired per recruiter has fallen by nearly half [2]. Two sides of the same hiring process, both underwater, neither trusting what the other side is telling them. That's the shape of the thing this investigation actually goes looking for.

In 1973, Michael Spence formalized something every hiring manager already knew in their bones: an employer can't directly observe how good a candidate actually is, so hiring runs on signals instead — and a signal only carries real information if it costs something to produce [6]. A college degree worked as a signal not because the coursework taught job skills, but because finishing one took years most people wouldn't spend if they weren't capable. A polished résumé worked because writing one well took judgment. A fluent, specific answer in an interview worked because you genuinely couldn't fake command of your own experience for forty-five minutes, live, in front of another person.

Every hiring system eventually gets arbitraged — grade inflation dulled the transcript, cover-letter templates dulled the cover letter, interview-prep courses dulled the unrehearsed answer. What's different now is the speed and the completeness: generative AI drops the cost of producing hiring's core signals close to zero, on both sides of the table, more or less simultaneously. Spence's own framework predicts what should happen next — when a signal goes from costly to free, the market stops trusting it and hunts for one that still costs something. Two recent studies actually went and checked whether that's really happening.

Start with the counterintuitive result. Economists at NBER randomized algorithmic writing assistance — grammar, style, spelling — across 480,948 real jobseekers on a major online labor market and found that the assisted group was hired 8 percent more often, with no evidence that employers ended up less satisfied with who they hired. The effect was largest for non-native English writers, who made up the bulk of the sample [4]. Cleaner writing, it turns out, doesn't just fake a signal — for a lot of applicants, it helps a real one come through that bad prose was previously hiding.

Then a second, more targeted study went straight at Spence's prediction. Researchers tracked the rollout of an AI cover-letter tool on a large freelance platform and found that access to it did raise the textual match between cover letters and job posts, and did raise callback rates. But the relationship between that match and getting a callback — the signal's actual predictive value — fell by 51 percent after the tool arrived. Employers didn't shrug; they adapted, shifting weight onto a signal AI can't fabricate: a candidate's actual prior work history [5]. That is Spence's arbitrage cycle, measured in real hiring data, not asserted in theory.

On the employer side of the same transaction, the applicant-tracking systems meant to manage the flood have their own well-documented failure mode. A Harvard Business School study conducted with Accenture found roughly 27 million Americans who are actively job-searching, genuinely qualified for roles they're applying to, and systematically filtered out before a human ever sees the application — mostly because of rigid keyword matching and parsing errors that silently discard a resume's actual content. In that same research, 88 percent of surveyed employers believed their own system was screening out good candidates [3]. And there's a second, older mechanism working against candidates from the other direction: peer-reviewed research on job postings during and after the last major downturn found that a one-point rise in unemployment pushed up degree requirements in postings by nearly half a point — and that roughly 40 percent of that tightening never reversed once the labor market improved. Employers ask for more when they can, and the asking mostly sticks [7].

Where serious people disagree is what to make of these patterns. Are ghost jobs a symptom or a strategy? Competing hypothesis The academic evidence that they cluster among large, high-skill employers cuts both ways [1]. One reading: big companies with sophisticated talent-acquisition teams are the most likely to deliberately farm applications for market intelligence and pipeline-building, because they can afford to. The other reading: those are exactly the employers running the highest volumes through the most overloaded recruiting teams, so what looks like strategy is really just a listing nobody got around to closing. Nobody has yet separated intent from neglect at scale. And does AI screening make selection better or worse? One side argues that a consistent algorithm, whatever its flaws, beats an exhausted recruiter skimming a two-hundredth résumé at 6 p.m. — and that structured, AI-assisted evaluation can widen a funnel rather than narrow it. The other side argues that a model trained on a company's past hiring simply launders yesterday's preferences into today's rejections, invisibly and at scale, and points out that the parties measuring adoption are rarely the ones measuring accuracy [3]. Independent, peer-reviewed evaluation of production screening systems barely exists — it remains the single largest evidence gap in this entire investigation.

Hiring doesn't break in isolation. The signal-cost collapse documented above [4][5] is interacting directly with a labor-market shift already underway: entry-level openings in AI-exposed occupations are contracting even as recruiting teams shrink and applicant volume climbs [2][9]. A companion investigation on this site (Are AI Layoffs a Strategic Mistake?) documents the layoff side of that same contraction — fewer junior hires upstream means fewer people with the verifiable work history this piece's evidence suggests employers are now falling back on [5]. The screening software meant to manage the resulting flood [3] is itself a second AI system sitting on top of the first, filtering the applicant pool the first system helped inflate.

Spence's signaling framework predicts the market hunts for a new costly signal once the old one goes cheap — but it doesn't say which one wins. Below are three candidates the evidence already gestures toward, not a forecast of which one hiring actually lands on.

Every mechanism in the modern hiring funnel — résumé, cover letter, unscripted interview answer — worked because it cost something real to produce, and generative AI made all of them cheap at once, on both sides of the table [5]. Which replacement signal ultimately wins — verification, negotiation, or raw experience — isn't settled by today's evidence, and each of the three paths above shares one uncomfortable feature: every one of them makes a first job harder to land than a fifth. Public, trackable indicators worth watching over the coming months and years:

• Hiring process signals — Whether the share of interview processes including an in-person round keeps rising — the cleanest early test of the verification turn [5]. Growth in paid-trial and work-sample hiring as a share of total hiring.
• Recruiting market data — Whether the estimated ghost-job rate moves as recruiting teams stabilize or keep shrinking [1][2]. Whether application volume eventually settles down, the fingerprint of a negotiation equilibrium taking hold.
• Screening technology oversight — The first independent, non-vendor accuracy evaluations of production AI screening systems — still the field's biggest open question. Whether major employers start formally disclosing permitted AI use rather than policing it.
• Long-term labor market fingerprint — Whether skills-based hiring's practice finally catches its own pronouncements, or the experience re-anchor shows up in the data instead [8]. Whether verification costs come to dominate hiring budgets, or entry-level hire rates keep diverging by prior experience — each path leaves a different signature.

If you're the one hiring, and you could keep exactly one signal — a résumé, a work sample, an in-person conversation, or a trusted referral — which would you keep, and what does that say about where you think hiring is actually headed? Tell us. We read every message.