The job market for lemons

In 1970 a young economist, George Akerlof, published a paper called ‘The Market for Lemons’. Using the second-hand car market as an example, he showed how asymmetric information about quality can cause market collapse. Sellers know if their car is a ‘lemon’; buyers do not. The price settles somewhere between the value of a lemon and a non-lemon. Non-lemon sellers then withdraw, and soon only lemons remain. Though simple, the principle won Akerlof a Nobel Prize in 2001. In this essay I want to use his work to explain how the job application market in 2026 has suffered its own version of this collapse.

​

The job market as a matching problem

A job market is not like a traditional market. There is no price as such. Instead it is a system for matching applicants to jobs. An employer posts a vacancy and wants to find a candidate who is two things: qualified and serious. Qualified in that they have the right skills and, in some cases, formal credentials. Serious in that they are genuinely interested in the role. Assessing qualifications has always been hard. Employers have built a battery of tests and interviews to do so, typically arranged as a funnel: cheaper assessments first, costlier ones like in-person interviews later. None are perfect and all incur cost. Seriousness is different. If an employer is going to move an applicant through to the expensive later stages, they want to know the candidate actually wants the job. Anyone who has hired people knows the frustration of an exciting candidate who pulls out at the offer stage, or worse, accepts and then reneges before the start date. But historically this was not a widespread problem for a simple reason: applying for a job took time, and that time was effectively a cost that filtered out the unserious. My first job application was for a Saturday shelf-stacking role at the local Co-Op. I completed a paper form and dropped it off in person. My next was an investment banking internship: an online application with a CV, cover letter, and several free-text questions. It took three or four hours per application. I applied for half a dozen and was serious about every one.

​

What changed

Two things happened in sequence. First, one-click apply. LinkedIn and other jobs boards allowed applicants to use their profiles to apply with a single click. Some were undoubtedly motivated by the Cost-per-Click’ model. Others saw applications as an onboarding flow and sought to maximise conversation. This reduced the cost of applying and unsurprisingly increased the number of applicants per role. Employers could infer less about an applicant’s interest from the mere fact of applying. It made filtering harder, but it was manageable: employers could require cover letters or add other low-cost screening stages to signal seriousness. Then came AI. Applicants could now tailor every application to the job description and generate professional cover letters at a click. Soon tools were built on top of LLMs to automate the entire process: zero-click apply with tailoring. Combined with a tough job market, the number of applications per role skyrocketed. Workday reported that the volume of applications on its platform grew nearly four times faster than vacancies in 2024. According to Greenhouse, the average number of applications per job has doubled since the release of ChatGPT.

​

The collapse

Now picture an employer who posts a job and receives a thousand applications. Perhaps a hundred of those applicants are serious: they have researched the company, understood the role, created a bespoke CV, and written a thoughtful cover letter. More to the point, they actually want the job. The other nine hundred have used an AI tool to apply for this and hundreds of other jobs. They may not even know they have applied - several recruiters have reported calling up applicants who have no idea what they have applied for. Why would people submit an application they are not serious about, even if it costs nothing? Because it is easier to apply for everything than to spend time reviewing postings and deciding which ones are interesting. It is not that the unserious applicant has zero interest, just that the probability of them being genuinely interested is a particular role very low: it’s option value. The employer would happily spend real time reviewing the serious hundred: reading their CVs, considering their cover letters, providing feedback. But they cannot do that for a thousand. So they resort to blunter instruments: discarding half the pile, running an AI screening tool to surface the ‘best’ CVs, or falling back on nepotism and only interviewing people who already know someone at the company. Serious applicants soon become aware of this. Scroll through any jobseeker subreddit and you will find a rich discourse about how companies actually review applications. The result is predictable: serious applicants are no longer prepared to spend hours on considered applications. They too resort to spray and pray, or give up entirely. Just as Gresham’s Law tells us that bad money drives out good - because people hoard sound currency and spend debased currency - unserious applications have driven out the serious ones.

​

A double-sided problem

You can take this further and describe it as a double-sided Akerlof problem. In the used car market there is one hidden variable: the seller knows whether the car is a lemon, but the buyer does not. In the job market there are two. On the applicant side, seriousness is hidden. The employer cannot tell whether an applicant has spent an hour researching the company and crafting a thoughtful application, or whether an AI tool fired off a generic one while they slept. This is the direct analogue of Akerlof’s lemon: the applicant knows, the employer does not. But there is a second asymmetry running in the opposite direction. The employer knows how much effort they will invest in reviewing each application: whether it will get ten minutes of careful reading or ten seconds of automated screening. But the applicant does not. An applicant deciding whether to invest time in a serious application is making a bet on the quality of the review it will receive, with no way of knowing the odds. Each side’s hidden information poisons the other’s incentives. If applicants suspect their carefully written cover letter will be discarded by an AI filter or buried under 999 others, they stop writing careful cover letters. If employers suspect that most applications are AI-generated spam, they stop reading them properly. Both responses are individually rational. The result is a equilibrium in which all applicants submit passive, untailored applications and all employers resort to cursory, arbitrary reviews. The serious applicant and the diligent employer both exist, but neither can find the other. Everyone loses.

​

How does this get solved?

Akerlof’s original paper identifies several mechanisms that can rescue a market from information collapse: independent quality assurance, brands that raise the cost of dishonesty, and licensing. These map well onto the problem of verifying qualifications, which is what certifications and credentials already do. Employers are rightly investing more here as AI-enhanced CVs become ever more creative. But seriousness is the harder problem, and the solutions on offer range from crude to transformative. In the short term employers drowning in applications are reaching for a quick fix. We see three main sticking plasters being used: deliberate friction, nepotism and AI vs AI. Nepotism. The quickest way to know an applicant is serious is if someone you trust tells you so. This is already the dominant mechanism in many professional industries: if someone at the company vouches for you, your application gets a proper look. A few startups are trying to productise this through referral networks and warm-introduction platforms. But as a primary matching mechanism it is obviously unfair, entrenching advantage among those who already have the right connections. Deliberate friction. The antithesis of one-click apply. By making it harder to apply - requiring bespoke exercises, lengthy free-text responses, or even handwritten postal applications - employers can ensure that only serious candidates will persevere. A related approach is rationing: capping the number of applications each person can make, as UCAS does with university places or the American Economics Association does for academic economics jobs. These work, but they accept the basic architecture of the market and simply try to filter more effectively within it. They also penalise candidates who face genuine barriers to spending time on applications, such as those already working long hours or managing caring responsibilities. Employers might end up trading quality for seriousness. AI versus AI. The instinct of many employers has been to fight fire with fire: deploy AI screening tools to detect AI-generated applications. This is already widespread, but it merely produces a cat-and-mouse arms race. Each side optimises against the other. Nobody wins, and the serious applicant who wrote their own cover letter may well be the collateral damage. It also opens up serious ethical and legal questions around automated decision making - watch this space. These are all short-term patches. They treat the symptom - too many unserious applications - rather than the cause, which is that the basic architecture of the market is broken. Replacing the job post. The more promising direction is to change the structure of the market itself. Rather than employers posting vacancies and waiting for a thousand applications to roll in, they go looking for the candidates they want. This is not new in principle - headhunters and executive recruiters have always worked this way - but AI makes it viable at a completely different scale and price point. If a company can describe what it needs and an AI agent can search a pool of candidate profiles to find credible matches, the entire application bottleneck disappears. The employer gets a shortlist of people who are plausibly qualified; the candidate gets approached about roles that are plausibly relevant. Both sides signal seriousness through specificity rather than effort. This is the hypothesis behind Rodeo, the careers platform we are building: an AI agent that knows what a candidate can do and proactively matches them to opportunities, replacing the spray-and-pray application with something closer to a mutual introduction. There is a second structural shift that may reinforce this. In some industries, AI is making it easier to disaggregate jobs into discrete tasks. Briefing workers, measuring output, and ensuring knowledge transfer are now tasks that can be at least partially automated. I have written elsewhere about this potential gigification of the labour market. If work moves in this direction, the traditional job post becomes even less relevant. How we match up people and work is a fundamental question of both our economy and society. How well we do it has implications for unemployment, productivity and ultimately, our happiness. AI, by accelerating the turnover of jobs, has made this even more pressing. But it has also triggered a genuine Akerlof-style failure in the jobs market. Both sides have rational incentives that, in aggregate, produce a worse outcome for everyone. The short-term fixes are knee-jerk reactions. The longer-term answer is not to keep filtering within a broken system but to replace the system’s basic architecture: from applicants chasing job posts to platforms that do the matching for both sides. Until that shift happens, expect more lemons.