Photo by Hugo Jehanne on Unsplash

Each time my own company has posted a role even loosely related to data systems, reporting, or BI architecture in the last couple of years, I know about it. Not because I receive an internal announcement, and not because someone forwards me the listing.

I know because, within a day or two, I receive a flood of near-identical interest emails; variations on a template, addressed to me by name, referencing my role, and asking whether I might be the right person to discuss the opening.

I worked out long ago that this is part of an AI-driven outreach service. It locates new postings, scans LinkedIn for employees in adjacent functions, and helps automate proactive contact with the people most likely to be close to the hiring manager. The logic is sound; if you cannot trust the front door of a job posting to actually lead anywhere, you walk around to the side and knock on a window.

More recently, I have noticed something else. Interview responses have started to circle questions in a particular way – a structure that builds out from a definition, lists adjacent considerations, qualifies the answer with a caveat, and usually lands somewhere safe.

The shape is familiar because I see it every day in another context. I have started, half-seriously, asking how a frontier model would respond to my interview questions, and using that as a benchmark for what to watch for in a human interview.

When the answers are too similar, the question stops being whether the candidate used AI and becomes whether the interview still measures what we think it measures.

These are surface observations from my own limited context. But the surface nearly always describes the structure it covers.

With the next round of Job Openings and Labor Turnover data scheduled for release on May 5, this seems a good moment to look at what the surface is sitting on top of.

What the data shows

In February 2026, the U.S. Bureau of Labor Statistics reported 6.9 million job openings and 4.8 million hires. The hires rate, 3.1% of total employment, was the lowest since April 2020, the early months of the pandemic shock. Quits stood at 3.0 million, a rate of 1.9%, falling well below the peak of around 3.0% reached in 2021 and 2022. Layoffs and discharges were 1.7 million, a rate of 1.1%, near historical lows.

These numbers do not describe a collapsing labor market, but they don’t describe a booming one either. They describe a market largely frozen in place.

Openings are elevated relative to hires. The ratio sits near 1.44 openings per hire, well above the 1.0 to 1.2 range that held through most of the pre-pandemic decade. Quits have collapsed, which means workers are not voluntarily moving as often. Layoffs are contained, which also means employers are not involuntarily moving them. Still, hires are at a five-year low outside acute disruption.

Visible activity at the surface in the form of postings, applications, interviews, and screens, it seems, has decoupled from the underlying mechanism the market exists to perform, which is matching workers to work.

A frozen market is one where everyone keeps performing the motions, and very little actually moves.

The structural argument of this piece is that the freeze is not the result of any single actor behaving badly. Employers are not, in the main, lying. Similarly, workers are not overall unwilling. The freeze is the result of a signaling layer that has been asked to carry more weight, in more directions, than the artifacts of the labor market can bear, and of an automation layer that has lowered the cost of generating those signals on every side simultaneously.

What follows is an attempt to look at each layer in turn: the posting, the application, and the role itself.

The posting

A job posting used to imply that a job existed, though the implication was never perfect. There were always postings that lingered after a hire, postings opened to satisfy an internal requirement before an external candidate was selected, and postings used to gauge a salary band. But these were edge cases.

The default reading was that a posting indicated an opening, and the indication was usually correct enough to be useful.

Now, though, a posting can mean active hiring. But it can also mean pipeline building, or benchmarking. It can mean procedural compliance, just as easily as it can mean investor, customer, or competitor signaling. It can even mean pressure on current employees, or nothing more than administrative residue.

These are not morally or operationally equivalent. But from the outside, they wear the same uniform.

This is where the first AI layer enters. In many companies, applicant tracking systems and AI-driven screeners now act as a prefilter, often before any human reviews a resume. The cost of receiving applications has fallen dramatically, as has the cost of processing them. When the cost of receiving and processing applications falls, the cost-benefit calculation around posting a role shifts. Posting was once a meaningful commitment of recruiter time. It is now closer to free.

When the cost of generating a signal falls, the equilibrium quantity of that signal rises, regardless of the underlying intent. This is one plausible structural reason openings and hires can drift apart. The per-posting hiring intensity has dropped, because the marginal cost of an additional posting – and the marginal cost of an additional purpose attached to that posting – is now low enough that the artifact has been asked to do work it was not designed to carry.

The result is that a posting is no longer a signal of a job. It is a signal of something, certainly, but the receiver cannot tell which something, and increasingly cannot tell whether the something is anything at all.

The application

The applicant side of the market has done exactly what game theory predicts. If postings are noisy, screening is automated, and the cost of applying is low, the rational response is to apply to more postings. If you want to be efficient in doing so, you’ll use automation to do it.

AI-powered application tools now auto-fill forms across multiple platforms. Resume optimization services rewrite a single CV into dozens of variants, each tuned to the keyword profile of a specific posting. Cover letters are generated, customized, and submitted at a pace no individual writing them by hand could match. Interview support tools coach candidates in real time, generate answers to behavioral questions, and run mock interviews with AI evaluators. The shape of those answers, the structure I now recognize across actual interviews, is the natural output of a system optimized to produce competent, defensible, low-risk responses to a known distribution of questions.

This is not, in itself, a moral failure. The applicant is responding to a market in which the signal value of a single application has been driven down by the screening environment on the other side. If a recruiter sees five hundred applications for one role and uses an AI screen to narrow them to twenty, the rational applicant strategy is not to write one careful application for one role, but to generate enough applications, tuned closely enough to the screening criteria, that some inevitably pass through. The applicant cannot unilaterally disarm, as doing so would effectively remove them from the pool.

The result on the employer side is symmetrical. Recruiters drown in low-signal applications. They tighten the screen. Applicants respond with better automation. Both sides escalate. The arms race is rational at every step, and at every step it destroys a little more of the signaling content of the underlying artifact.

There is a clean structural observation embedded in this: The application used to be a costly signal. Writing a careful cover letter, tailoring a resume, and researching the company all took time, and the willingness to spend that time was itself information. It told the employer something about the applicant’s interest, attention, and judgment. When the cost of producing those artifacts collapses, so does their information content. The application is no longer a signal of interest. It is a low-cost option exercised against a low-cost screen.

The JOLTS quits data sits on top of this layer. Quits at 1.9%, now well below the 3.0% peak, tell us that the voluntary mobility channel that historically disciplined employers and rewarded workers, has narrowed. Workers can see the market from the outside. They are looking at it through the same fog of ghosted applications, automated screens, and interview processes that evaporate after two conversations.

The rational response is to stay where they are. The market has frozen them in place.

The role

The third layer is the job itself, once you have it. Here too, the signaling content has degraded.

ADP’s National Employment Report for March 2026 shows job-stayers receiving 4.5% year-over-year pay growth. Job-changers, those who switched employers within the prior twelve months, received 6.6%. The gap is roughly 2.1 percentage points. That gap was over 7 percentage points at its 2022 peak, so while it has compressed sharply, it has not closed.

The implicit message from this gap is precise: Loyalty is mildly punished, while mobility is mildly rewarded. Neither lever is strong enough to discipline the system. A worker who stays gets pay increases that approximate inflation. A worker who leaves gets pay increases that meaningfully exceed it, but only by a margin that has shrunk to a multi-year low.

The old bargain “if loyalty is not rewarded, mobility will be” still holds, but more weakly than it has in some time.

Internal promotion has thinned alongside this, as promoting an internal candidate often creates a backfill chain: the open seat must be filled, often by another internal candidate, whose seat must then be filled, and so on, until the chain bottoms out at an external hire. Each link in that chain is operationally expensive, while external hiring at the original level avoids the chain entirely. Even where internal promotion remains a stated value, the mechanics often work against it.

The career-path signal that staying produces advancement degrades from the accumulated weight of many locally rational choices.

And then AI arrives at the role itself. Workers in roles increasingly exposed to language model capabilities – analytics, reporting, content production, software development, and customer support – face a question that was not on the table five years ago: ‘Which parts of my function will be automated, on what timeline, and is my employer investing in me or quietly preparing to replace me?”

ADP’s Today at Work 2026 report found that only 22% of global workers strongly agreed their job was safe from elimination. Four out of five do not share that confidence. This compounds the freeze.

Workers who feel insecure in their current role might historically have responded by moving; testing the market, accepting an offer elsewhere, and capturing the mobility premium. But the application side is now a fog of automated screens and ghosted processes, and the premium itself has compressed. The exit is harder, the reward is smaller, and the role they are exiting feels less stable than it did. The rational response is to keep one eye on the door and the other eye on whatever signals leadership is sending about AI investment, headcount targets, and quarterly margin pressure.

They watch carefully, and often they do not move. They appear, on the surface, to be loyal.

This is the spiral. Again, it does not require any actor to behave badly. Employers optimize for cost, optionality, and operational simplicity. Workers optimize for self-protection, mobility, and signal interpretation. Each side’s rational response makes the other side’s environment slightly worse, and each side’s automation reduces the cost of escalating its own response further.

The system teaches both sides to distrust each other, and then treats the distrust as evidence that the other side was never worth trusting.

The frozen market

Pull back to the surface. The U.S. labor market in early 2026 shows 6.9 million open positions and the lowest hires rate since the early pandemic. It shows quits well below their recent peak, layoffs near historical lows, and a job-changer pay premium that has compressed to a fraction of what it was just a few years ago. By these measures, the market is not collapsing, but it is also not matching.

The signaling layer has degraded in a coordinated way:

• A posting no longer reliably signals a job.
• An application no longer reliably signals interest.
• A screen no longer reliably signals judgment.
• A retained role no longer reliably signals security.
• A promotion path no longer reliably signals advancement.
• A compensation package no longer reliably signals investment.

When all of these signals degrade simultaneously, the market freezes, and people stay where they are, while companies post without hiring and applicants apply without expecting. The only new factor lies in how AI fills the gap on both sides, generating motion without movement.

The role of automation in this is worth stating carefully. AI did not create the dysfunction; the artifacts of the labor market were already overloaded. What AI did do, however, was lower the cost of generating signals on every side of the market at once, which accelerated every existing incentive misalignment toward its limit. The equilibrium did not change in kind, but in degree. However, the degree was enough to make the market stop functioning as a matching system in any reliable sense.

The deeper observation, and the one that connects this piece to a broader pattern, is that trust is labor-market infrastructure. The matching function depends on a layer of signals that the participants can read with reasonable confidence. When that layer degrades, the entire system pays a coordination tax. Recruiters cannot trust applications, so they tighten screens. Applicants cannot trust postings, so they apply more broadly and rely on outreach automation to bypass the front door. Workers cannot trust internal advancement, so they keep one eye on the exit. Firms cannot trust retention, so they invest less in development.

Each rational adjustment makes the next round of trust harder to extend.

Worse, no single actor in this system has an incentive to unilaterally restore the signals. No employer benefits from being the only one to remove AI screening; they would just pay more for the same inputs. No applicant benefits from being the only one to apply by hand; they would be drowned out by the volume of people who use AI services. Similarly, no firm benefits from being the only one to commit to internal promotion when external hiring is operationally cheaper.

The equilibrium is self-reinforcing. Each participant is behaving rationally inside a structure that rewards behaviors that, taken together, destroy the system’s capacity to do the thing the system exists to do.

What we are watching, when we look at the JOLTS series, ADP wage data, and the surface phenomena of ghost jobs, automated outreach, and AI-shaped interview answers, is the consequence of incentives that have stopped translating across the participants in the labor market. The artifacts remain familiar, and the activity continues. For all common measures, the motion is real, but the matching, increasingly, is not.

The next JOLTS release will arrive on May 5. It will report new numbers for openings, hires, quits, and separations.

Again, we will see the surface move. The question is whether the market beneath it is matching again, or whether we are only watching another month of signals generated by a system that no longer trusts its own language.