I've been looking at what generative AI is doing to recruitment, and there is one part of it I think we're underestimating.

Most of the discussion is about efficiency, cheating, fake CVs or whether AI will eventually replace recruiters. Those things matter, but they seem secondary to a more fundamental change happening underneath them.

We are changing the economics of the signals we use to judge people.

And hiring happens to be a very good place to see what that does to a system.

291 applications for one hire

Ashby recently analysed more than 109 million applications across 247,000 jobs.

Applications per hire roughly tripled between 2021 and 2024. Throughout 2025, employers received more than 300 applications for every person they eventually hired. The latest average sits around 291.

There are plenty of reasons for that increase. Remote work widened the potential candidate pool. Easy Apply removed friction. The labour market changed. Job aggregation improved. People simply became able to apply to far more jobs than they could before.

Generative AI pours fuel onto an already existing dynamic because it radically reduces the effort required to create a credible application.

A tailored cover letter used to take some time. Adjusting your CV to a vacancy took some interpretation. Writing a good motivation required at least a reasonable ability to understand the role and explain why your experience matched it.

Those were never particularly good measures of whether someone could actually do the job. But they did contain some information about the person producing them.

Now I can give ChatGPT my CV and a vacancy and have a perfectly reasonable application a few seconds later.

Do that for 50 vacancies and the marginal effort gets pretty close to zero.

The strange part is that my capability didn't change at all.

My representation of capability did.

The signal is starting to inflate

This is the part I find interesting.

Lighthouse Research and Criteria surveyed 998 employers this year. Only 33% said they were very confident that resumes accurately reflected candidates' real skills. At the same time, most organisations still use the resume somewhere near the beginning of the selection process.

That leaves hiring with a pretty uncomfortable contradiction.

We still depend heavily on an artefact that the people making the decision increasingly don't trust.

Generative AI didn't suddenly make CVs useless. CVs were imperfect signals long before ChatGPT existed. People exaggerated experience, hired professional writers and learned how to optimise for recruiters decades ago.

AI changes the scale and the economics of that behaviour.

A strong signal needs some characteristic that makes it useful for separating one person from another. It might be difficult to imitate, expensive to acquire, independently verifiable, scarce, directly connected to the underlying capability, or produced under conditions where gaming it is difficult.

If AI makes the outward appearance of a signal cheap enough for almost everyone to reproduce, some of that separating power disappears.

The strong candidate can produce a beautifully tailored application.

So can the average candidate.

Increasingly, so can someone who barely understands the vacancy.

The distance between their underlying capabilities may still be enormous. The observable distance between their applications becomes much smaller.

I think signal inflation is a useful way to think about that.

The employer then has to find the information somewhere else

Employers obviously don't just shrug and start hiring randomly.

They adapt.

And interestingly, Ashby's data suggests recruiters have actually become quite good at this once candidates make it deeper into the funnel. Despite the massive increase in applications, candidates reaching interviews are converting into offers at rates above 2021 levels. Recruiter productivity has also recovered.

So I wouldn't describe this as hiring simply breaking down.

The pressure seems to be moving toward the point where hundreds of candidates need to be reduced to the handful worth investigating properly.

Criteria's employer research gives a clue as to where organisations are looking instead. 98% of the surveyed employers considered assessments, structured interviews and work samples more trustworthy than resumes. In a separate survey of more than 2,500 candidates, 68% were open to abandoning the traditional resume in favour of processes built around other signals.

That makes intuitive sense.

When describing your ability becomes easier, organisations place more value on watching you demonstrate it.

But there is an annoying systems problem here.

Deep verification is expensive.

You can receive 300 CVs. You cannot realistically conduct 300 structured interviews, work simulations and supervised assessments.

So while the top of the funnel becomes increasingly cheap to enter, the amount of serious attention available further down the funnel remains scarce.

That makes the real question increasingly interesting:

Who gets access to the part of the hiring process where capability is actually observed?

Trust becomes more valuable when information becomes noisy

There is another possible adaptation appearing.

The Financial Times reported this week on recruiters increasingly revisiting referrals, recommendations and existing relationships as AI-generated applications become harder to distinguish. Someone recommending another person puts a piece of their own reputation behind that recommendation, which makes the signal harder to manufacture at scale.

I would be careful claiming that referrals are already taking over recruitment because of AI. The wider data doesn't establish that. Ashby's own dataset actually shows referrals shrinking as a proportion of total applications while inbound applications exploded.

But the incentive is worth paying attention to.

If publicly available signals become less informative, employers will naturally search for signals that are harder to fake. Reputation and trusted relationships are obvious candidates.

That solves one problem and can easily create another.

A referral tells me something useful because I trust the person providing it. But access to that signal depends partly on already being connected to someone whose opinion carries weight.

Someone with ten years inside an industry has plenty of that.

A graduate trying to enter it doesn't.

Neither does the immigrant whose professional network exists somewhere else, the career switcher or the genuinely brilliant person who happens to be terrible at networking.

We could therefore improve trust while accidentally reducing access.

And everyone involved can still be behaving perfectly rationally.

That is where the loop gets uncomfortable

Put yourself in the position of a candidate.

You know a vacancy might receive hundreds of applications. You have no idea whether a human will ever read yours, you know some form of screening is likely happening, and AI can tailor your application for almost no effort.

Using AI is an entirely logical response.

Now sit on the employer side.

Hundreds of increasingly polished applications arrive. Manually evaluating all of them isn't feasible, so you automate more of the initial filtering, add screening questions or move verification deeper into the process.

Also logical.

Candidates learn what the screening systems reward and optimise their applications accordingly.

Still logical.

Employers trust the resulting applications even less and add more verification.

Candidates now face increasingly long application processes while the chance of any individual application succeeding remains small, which makes spending hours carefully writing one application less attractive.

So they apply more broadly.

You can see where this goes.

Nobody necessarily caused the problem by behaving irrationally. Each participant responded reasonably to the incentives sitting directly in front of them.

Collectively, they degrade the information architecture they all depend on.

That's a much harder problem to solve.

There is no CEO of the labour market

Employers can redesign their selection process. LinkedIn and other platforms can change how applications work. AI companies can change what their tools allow. Governments can regulate automated hiring. Universities can rethink credentials. Candidates can find new ways of demonstrating ability.

None of them controls the system.

Their incentives aren't completely aligned either.

Candidates want applying to become easier and faster. Employers want enough friction or verification for an application to carry information.

Platforms generally benefit when people apply and interact more. Recruiters don't necessarily benefit from receiving more applications.

Employers want verification without building an enormously expensive hiring process. Candidates reasonably don't want six interviews, three tests and biometric identity checks just to prove they're a real human who knows Excel.

Regulators want transparent and fair processes. Organisations still need to fill vacancies quickly.

There isn't one obvious desired state everyone can optimise toward.

Which probably means we're going to watch the hiring market adapt through hundreds of smaller changes rather than one clean solution.

And I suspect the direction of that adaptation matters a lot.

Proof may become the scarce thing

AI is making good language abundant.

It is making polish abundant.

Tailoring is becoming abundant.

Even reasonably convincing representations of knowledge and experience are becoming cheaper every month.

The interesting economic consequence is what becomes valuable because those things are abundant.

Observed performance.

Work produced under known conditions.

Verified experience.

A reputation somebody else is willing to stake.

Longitudinal evidence that you consistently know what you're doing.

Judgment demonstrated during an actual interaction.

In other words, credible proof of competence becomes more valuable as producing the appearance of competence becomes cheaper.

Hiring may simply be one of the first places where we're being forced to deal with this at scale.

I doubt it will be the last.

Education has the same problem when an essay stops reliably telling us what a student understands. Consulting has it when a polished analysis no longer tells a client how much thinking actually happened behind it. LinkedIn itself has it when anyone can publish ten perfectly articulated “thought leadership” posts per day.

We have spent a lot of time thinking about what happens when AI makes knowledge work cheaper to produce.

I think we should spend a little more time thinking about what happens to the systems that relied on the cost and difficulty of producing that work as evidence of capability.

Because once everyone can look competent, we still need some way of figuring out who actually is.