Doom Loop: Proven AI Job Market Risks to Avoid Now

Doom loop is the phrase Greenhouse chief executive Daniel Chait keeps reaching for when he describes what artificial intelligence has done to hiring, and WIRED built a whole feature around it on 4 September 2026. The mechanism is unusually simple for something this destructive. Candidates have a problem, so they point AI at it. Employers have a problem, so they point AI at it. Each side solves its own problem in a way that makes the other side’s problem worse, and the wheel turns again.

Chait’s own summary is the clearest description anyone has offered: “We’ve got this tragic situation where each side has a problem. They’re using AI to solve their own problem, but in ways that make the problem worse. And so more AI use begets more AI use, to no one’s benefit. The more it’s happening, the worse it gets.” That is not a complaint about lazy applicants or lazy recruiters. It is a description of a feedback system with no natural brake.

This article takes the doom loop apart mechanism by mechanism, puts the published numbers next to each other, and works out what a hiring team can actually do about it. If you follow AI stories that reshape how ordinary work gets done, they land in our AI models and tools hub as they happen. The short version for anyone recruiting right now: the volume is not going back down, so the only exit is to stop competing on volume.

Table of contents

What the AI Hiring Doom Loop Actually Is
The Numbers Behind the Job Market Doom Loop
Four Turns of the Doom Loop, Step by Step
Why the Doom Loop Destroys Trust on Both Sides
What the Doom Loop Costs an Employer
The Compliance Bill Hiding Inside the Doom Loop
Human Review Is the Only Exit From the Doom Loop
Signals That Still Survive the Doom Loop
A Practical Doom Loop Escape Plan for Hiring Teams
What Job Seekers Should Do Inside the Doom Loop
Where the Doom Loop Goes From Here
Doom Loop Questions Employers Keep Asking
References and Further Reading

What the AI Hiring Doom Loop Actually Is

The doom loop is not a metaphor for “AI is bad for jobs”. It is a specific, observable feedback cycle inside the application pipeline, and it has four moving parts.

The candidate side of the loop

Applying used to cost something. Writing a tailored covering letter took forty minutes, so most people applied to eight or ten roles they genuinely wanted. Generative AI took that cost to roughly zero. Around three in four job seekers now use AI to write or polish an application, and the rational response to a near-free action is to do far more of it. That is the first turn of the doom loop, and nobody in it is behaving irrationally.

The employer side of the loop

On the other side, a recruiter opens a requisition and finds hundreds of applications that read almost identically, because they were produced by the same handful of models working from the same job description. Human differentiation has been sanded off. So the recruiter feeds the pile into the AI ranking feature inside their applicant tracking system, because reading it by hand is no longer physically possible. That is the second turn of the doom loop.

Why the loop closes instead of settling

Here is the part that makes it a doom loop rather than a one-off adjustment. Candidates who get filtered out by an AI ranker do not conclude that they should apply more carefully. They conclude that the system is a lottery, so they buy more automation and apply to more roles. That pushes volume higher, which pushes employers further into automated ranking, which makes the process feel even more arbitrary. Each turn of the doom loop is a rational local decision that produces a worse global outcome.

The economic backdrop that tightens it

None of this happens in a vacuum. The doom loop is running inside a low-fire, low-hire economy where postings are scarce, scams and ghost jobs are common enough that several US states are drafting laws against them, and the ratio of applicants to openings is already brutal. Scarcity turns up the gain on every feedback path in the system.

Side of the loop	The problem they actually have	The AI they reach for	What it makes worse
Candidates	Almost no reply rate; the process is a black box	CV rewriting, cover-letter generation, auto-apply bots	Employers drown in near-identical applications
Employers	Hundreds of applications that look the same	ATS ranking, AI scoring, automated screening	Candidates conclude the process is arbitrary
Both	No reliable signal of genuine interest	More AI, applied harder, on both sides	Signal falls further, so the doom loop tightens
The market	Scarce postings, ghost jobs, low hiring rates	Nothing — AI does not create vacancies	Every feedback path above runs hotter

The Numbers Behind the Job Market Doom Loop

The doom loop is easy to describe and easy to dismiss as vibes, so it is worth putting the published figures in one place. Greenhouse sits underneath a large share of the technology sector’s hiring, which makes its platform data unusually direct evidence.

Volume figures from the platforms

In August 2026, Chait put it at 2,539 applicants for every ten open jobs across roughly 175,000 live roles on the platform. That is an average of 254 applicants per posting. Applications per recruiter, on the same platform, have risen 412%. LinkedIn has separately reported applications running around 11,000 per minute, up more than 45% year on year. Those three numbers describe the same doom loop from three different angles.

Behaviour figures from the candidate side

Greenhouse’s own candidate research puts 67% of US job seekers using AI somewhere in their search, 59% having altered a CV with it, and 45% of those admitting the alteration embellished their qualifications. Further down the same survey, 28% said they had used AI to create fake work samples and 22% said they had deployed automated bots to submit applications. Chait has also described candidates paying around $20 for tools that mass-apply on their behalf.

Trust figures, which are the ones that matter

The figures that actually explain why the doom loop persists are the trust numbers. Almost half of job seekers — 46% — say they have less confidence in hiring than they did a year ago, and 42% tie that drop directly to AI. Only 7% believe the market favours them. Meanwhile 72% report bait-and-switch job descriptions and 63% report being ghosted by an employer. A market where both sides expect bad faith is a market where the doom loop cannot self-correct.

Figure	Value	Source	What it tells you
Applicants per 10 open jobs	2,539	Greenhouse, Aug 2026	254 applicants per posting on average
Live jobs on the platform	~175,000	Greenhouse, Aug 2026	The sample is large, not anecdotal
Rise in applications per recruiter	+412%	Greenhouse, Aug 2026	Human capacity did not rise with it
Year-on-year rise in applications	+45%	LinkedIn	The doom loop is platform-wide, not vendor-specific
Job seekers using AI in the search	67%	Greenhouse candidate research	AI use is now the default, not the exception
Candidates running auto-apply bots	22%	Greenhouse candidate research	One in five applications may be unattended
Less confident in hiring than a year ago	46%	Greenhouse candidate research	Trust is falling, not just efficiency
Blame that drop directly on AI	42%	Greenhouse candidate research	Candidates can see the doom loop too

What candidates admit to doing with AI (%)

Used AI somewhere in the job search — 67%

Altered a CV using AI — 59%

Of those, embellished qualifications — 45%

Created fake work samples with AI — 28%

Deployed automated bots to apply — 22%

Four Turns of the Doom Loop, Step by Step

Watching the doom loop turn once is the fastest way to understand why adding more technology to either side cannot stop it.

Turn one: applying becomes free

A candidate who used to send ten applications a month sends a hundred. The marginal cost of the hundred-and-first is a few pence of inference. Nothing about this is dishonest, and telling people not to do it is telling them to compete with one hand tied.

Turn two: reading becomes impossible

The recruiter’s inbox goes from 40 applications per role to 254. At three minutes of genuine human attention per application, 254 applications is 762 minutes — 12.7 hours of reading for a single vacancy. No recruiter carrying eight open roles has 100 hours a week, so automated ranking stops being a choice and starts being an arithmetic necessity.

Turn three: the signal collapses

Automated ranking scores what it can measure. Underneath the marketing, most of it is natural language processing comparing one document against another, so what it can measure is surface similarity to the job description. Candidates work that out within one hiring season, and AI is extremely good at producing surface similarity. Within another season, the ranker is sorting on a feature every applicant can generate perfectly, which means it is sorting on noise.

Turn four: trust breaks and volume rises again

Candidates who cannot tell why they were rejected reasonably conclude that effort is not rewarded, so they revert to volume. Employers who cannot tell a genuine application from a generated one reasonably conclude that applications are not evidence, so they lean harder on the ranker. The doom loop has now completed a full rotation, with both sides worse off and both sides having behaved sensibly.

Reading 254 applications for one vacancy: hours of human attention

3 minutes each — 12.7 hours

60 seconds each — 4.2 hours

20 seconds each — 1.4 hours

6 seconds each — 0.4 hours

Why the Doom Loop Destroys Trust on Both Sides

Efficiency losses are annoying. Trust losses are structural, and they are what make this particular doom loop so hard to reverse.

Candidates stop believing the process is real

When 63% of candidates report being ghosted and 72% report job descriptions that did not match the actual role, a rejection carries no information. The applicant cannot distinguish “you were not right for this” from “a model scored you 6.2 and the cut-off was 6.5” from “the vacancy never existed”. Absent a signal, people assume the worst and behave accordingly, which feeds the doom loop directly.

Employers stop believing the application is evidence

The mirror image is just as corrosive. Once a hiring manager knows that a polished CV, a fluent covering letter and an articulate written exercise can all be produced in ninety seconds, none of those artefacts function as evidence any more. WIRED reported hiring teams noticing the tell in interviews — the pause, then the sound of typing, then a suspiciously well-organised answer.

The result is a market that punishes honesty

The uncomfortable equilibrium inside the doom loop is that a candidate who writes every application by hand is out-competed on volume, and an employer who reads every application by hand is out-competed on speed. Doing the right thing is individually costly for both parties. That is the signature of a genuine coordination failure rather than a moral failing.

What job seekers report about the market (%)

Hit bait-and-switch job descriptions — 72%

Were ghosted by an employer — 63%

Less confident in hiring than a year ago — 46%

Blame that loss of confidence on AI — 42%

Believe the market favours them — 7%

What the Doom Loop Costs an Employer

It is tempting to treat the doom loop as a candidate-experience problem and therefore somebody else’s budget. The costs land squarely on the employer, and most of them are invisible in the recruitment line of a P&L.

The cost of screening the wrong feature

If your ranker sorts on similarity to the job description, and every applicant can generate perfect similarity, then your shortlist is a random sample of applicants who own a subscription. You are paying full recruitment cost for a random draw, and the doom loop guarantees the randomness gets worse each cycle.

The cost of a longer, noisier funnel

Volume that carries no signal still costs money to move. Every one of those 254 applications consumes storage, screening time, scheduling effort and, increasingly, verification effort. Reports of employers reinstating in-person interviews purely to confirm that the candidate matches the application are a direct doom loop tax, paid in diary time.

The cost of the candidates you never see

The most expensive part of the doom loop is invisible by construction. Strong, employed, selective candidates apply rarely and give up quickly on processes that feel like a lottery. They are exactly the group most likely to opt out, and their absence never shows up in a report because they never entered the funnel.

The cost of hiring the wrong person confidently

When 45% of CV-alterers admit they embellished and 28% admit generating fake work samples, a screening process that cannot verify anything is a process that will occasionally hire someone who cannot do the job. That cost is not recruitment spend; it is a failed probation, a delayed project, and a team that has to absorb both.

The Compliance Bill Hiding Inside the Doom Loop

For a UK employer, the doom loop is not only a commercial problem. Automated candidate ranking is regulated processing, and the regulator has already looked at this market closely.

What the ICO actually found

Between August 2023 and May 2024 the Information Commissioner’s Office ran a series of consensual audits of AI recruitment tools and issued 296 recommendations and 42 advisory notes. It found tools inferring characteristics such as gender and ethnicity from a name, tools that let recruiters filter on protected characteristics, and tools collecting far more personal data than the purpose required and retaining it indefinitely without candidates knowing.

Why “the vendor handles it” is not an answer

The ICO’s remedy was aimed at buyers as much as builders: ask developers direct questions before you procure, insist on documented fairness and bias testing, and keep records of meaningful human review. If your applicant tracking system scores candidates, you are the controller for that processing, and the doom loop is not a defence. Our write-up of algorithmic auditing for HR sets out the checks in practical order.

Automated decision-making rules still apply

Ranking that materially determines whether a human ever sees an application starts to look like automated decision-making with significant effects, which carries its own transparency and human-review obligations under UK data protection law. The rules changed in 2025, and we covered the practical consequences in automated decision-making under the DUAA. Treating a ranking model as a neutral productivity feature is the most common mistake here.

Governance is cheaper than a complaint

None of this requires a large programme. It requires knowing which model scores your candidates, on what features, with what testing evidence, reviewed by whom, and retained for how long. That is the same discipline described in our AI governance framework for SMEs, applied to one system rather than an estate.

Human Review Is the Only Exit From the Doom Loop

The striking finding in WIRED’s reporting is that the split between employers who use automated ranking and those who refuse is not driven by company size or application volume. It is driven by philosophy.

Toshiba’s answer

Kim Jones, vice president of human resources at Toshiba, told WIRED plainly: “We have humans review every application.” She is untroubled by applicants using AI to polish their materials, noting it “is not really going to help with getting through the ATS”, because her screening turns on job requirements, salary expectations and whether the person is a rehire — facts, not prose quality. That is a doom loop exit built out of decisions no model made.

Why facts beat scores

Every criterion in that list shares one property: it is checkable and it is boring. AI cannot fabricate your salary expectation, your notice period, your right to work, or your rehire status. Screening on verifiable facts is immune to the generated-fluency problem that the doom loop is built on, because fluency is not one of the inputs.

The cost is real but bounded

Reviewing every application by hand sounds impossible at 254 per role until you separate mechanical filtering from judgement. Deterministic filters on hard requirements — location, right to work, salary band, licence, notice period — can cut a pile by most of its volume without any model scoring anybody. What remains is a human-sized shortlist, reached without ranking anyone by vibe.

Signals That Still Survive the Doom Loop

If generated text no longer carries information, the practical question for a hiring team is which signals still do. The test is simple: can a language model produce this artefact convincingly in under two minutes?

Signals AI can trivially produce

CV formatting, covering letters, enthusiasm, keyword coverage against a job description, written answers to generic competency questions, and polished LinkedIn summaries all fail the test. Weighting any of them is weighting noise, which is precisely how the doom loop converts effort into randomness.

Signals AI cannot fake for you

Verifiable employment history, a live conversation about a decision the candidate personally made, a piece of work you watch them produce, a reference who will take a phone call, and any hard credential you can check independently all survive. So does interest in your specific company, which is why Chait’s advice to candidates is to go beyond the headline employers.

Signals that survive but cost more

Structured work samples, paid trial tasks and technical exercises done live are the strongest survivors, and they are also the most expensive. The honest trade-off inside the doom loop is that reliable signal now costs more per candidate than it did in 2022, so you buy less of it and you buy it later in the funnel.

Screening signal	Can AI produce it in 2 minutes?	Survives the doom loop?	What it actually tells you
Covering letter quality	Yes	No	That the candidate has a subscription
Keyword match to the advert	Yes	No	That they pasted your advert into a prompt
Written competency answers	Yes	No	Very little, at any volume
Salary expectation and notice period	No	Yes	Whether a hire is even possible
Verifiable employment history	No	Yes	What they have actually done
Live conversation about their own decisions	No	Yes	Depth, judgement and honesty
Watched work sample or paid trial	No	Yes, strongest	Whether they can do the job
Reference who takes a phone call	No	Yes	How they behave on a team

A Practical Doom Loop Escape Plan for Hiring Teams

You cannot fix the market, and you cannot make candidates stop using AI. You can change what your own process rewards, which takes you out of the worst of the doom loop without waiting for anyone else.

First 30 days: stop scoring what AI writes

Turn off, or stop acting on, any ranking feature that scores free text. Replace it with deterministic filters on hard requirements you would defend out loud. Write down which model, if any, still touches a candidate record, and what it does. This step costs nothing and removes the largest source of doom loop randomness in most pipelines.

Days 30 to 60: rebuild the advert and the reply

Publish a genuinely specific advert — the real salary band, the real location policy, the real first six months of the job — because vague adverts are what make mass-applying rational. Then commit to a rejection that goes out within a fixed number of days to every applicant. Ghosting is a major input to the doom loop and it is entirely within your control.

Days 60 to 90: buy signal where it counts

Move your evaluation budget later in the funnel. One structured, paid, watched work task at shortlist stage buys more information than any amount of upstream scoring, and it is the part of the process the doom loop cannot degrade. Keep the shortlist small enough that a human decides, and record who decided and why.

Ongoing: measure the things the loop hides

Track time-to-first-human-response, the share of applicants who receive an outcome, offer-acceptance rate, and how many hires came from channels other than the open advert. Those four numbers move when the doom loop is hurting you, and none of them appear on a standard ATS dashboard.

Phase	Action	Owner	Evidence to keep
Days 0–30	Disable free-text ranking; document every model touching candidate data	Talent lead + DPO	System inventory and processing record
Days 0–30	Replace scoring with deterministic hard-requirement filters	Hiring manager	Written filter criteria per requisition
Days 30–60	Rewrite adverts with real salary, location and first-90-days detail	Hiring manager	Advert version history
Days 30–60	Guarantee an outcome to every applicant within a fixed window	Talent lead	Response-time report
Days 60–90	Introduce one paid, watched work task at shortlist stage	Hiring manager	Task brief and scoring rubric
Days 60–90	Record the named human who made each shortlist decision	Talent lead	Decision log
Ongoing	Report time-to-first-response, outcome rate, offer acceptance, source mix	Talent lead	Quarterly hiring metrics pack

What Job Seekers Should Do Inside the Doom Loop

The advice that follows from the mechanism is unintuitive, because the doom loop rewards the opposite of what feels productive.

Volume is the trap, not the strategy

If 254 people are applying to a posting and most of them used the same model to write the same letter, the hundredth identical application is worth nothing. Chait’s own advice is to go where the crowd is not: “Don’t just apply to, you know, OpenAI, because they’re in the news every day.” Think of the work you want and find the less famous companies doing it.

Aim at the part of the process a model cannot enter

Every signal that survives the doom loop is one a machine cannot produce on your behalf — a person who will vouch for you, a specific decision you made and can talk through, a piece of work you can be watched producing. Reallocating an hour from sending twenty more applications to securing one warm introduction is a straight upgrade in expected value.

Use AI, but not as a volume machine

Nobody sensible is arguing that candidates should stop using AI. Use it to research the company properly, to rehearse difficult questions, to tidy your own writing. The failure mode inside the doom loop is delegating the judgement about where to apply, because that judgement is the only thing separating you from the other 253 applicants.

Where the Doom Loop Goes From Here

Nothing about the current trajectory is stable, and there are only a few plausible ways this resolves.

Verification moves offline

The clearest early signal is employers reinstating in-person stages purely to confirm identity and competence. It is expensive and slightly absurd, but it works, and the doom loop makes it rational. Expect more of it, particularly for roles where a bad hire is costly.

Platforms start charging for volume

The economics of the doom loop change the moment applying stops being free. Rate limits, verified-applicant tiers and paid application credits are all being discussed, and each of them would dampen the first turn of the loop. Each also carries an obvious fairness problem, which is why nobody has shipped one at scale.

Regulators arrive before the market settles

Ghost-job legislation, the ICO’s recruitment audit work and the tightening rules on automated decision-making all point the same way. Employers who wait for the doom loop to sort itself out are likely to meet a compliance requirement first. The organisations who documented their screening logic early will find that unremarkable.

Or nothing changes and the loop keeps turning

The least dramatic outcome is also the most likely one in the short term: volume keeps climbing, ranking keeps degrading, and both sides keep escalating. That is what an infinite doom loop means. It does not end on its own, which is exactly why individual employers changing their own process is the only lever available right now.

Doom Loop Questions Employers Keep Asking

Is the doom loop just a Greenhouse marketing line?

The phrase is Chait’s, and Greenhouse does sell hiring software, so the framing is not disinterested. The underlying numbers, though, come from platform data and candidate surveys that line up with independent reporting from LinkedIn and with what recruiters describe unprompted. The mechanism holds up even if you discount the branding.

Should we switch off AI screening entirely?

Not necessarily — switch off AI scoring of free text. Deterministic filters, duplicate detection, scheduling automation and note-taking are all fine and none of them feed the doom loop. The dangerous category is anything that produces a rank order from prose a model wrote.

How do we spot an AI-written application?

You largely cannot, reliably, and building your process around detection is a losing strategy. Detection tools produce false positives on non-native English speakers and on anyone who writes formally. Design a process where it does not matter what wrote the application, because the decisive evidence comes later.

Does any of this apply to a small team hiring one person?

More than it does to a large one. A ten-person company has no recruitment operations function to absorb 254 applications, so the temptation to lean on whatever ranking feature the ATS offers is strongest exactly where the governance is weakest. The 30-day steps above are the ones that matter most at that size.

What is the single highest-return change?

Publish the real salary band. It cuts application volume from people who would never accept the role, raises trust with the ones who would, and it is the one change that reduces the doom loop’s input and improves candidate experience at the same time.