Why a Good Application No Longer Means a Good Candidate

In 2022, the average open role at a company using Greenhouse drew 116 applications. By 2025 it drew 244. Over the same period the average recruiting team shrank from ten people to under five, so the applications each recruiter handled rose from 146 a year to 746.

Those numbers describe a workload problem. Underneath sits a harder one. When every application arrives well written, closely matched to the posting and free of the small errors that used to separate careful candidates from careless ones, what is a recruiter actually reading?

What a résumé was doing when it worked

Michael Spence won a Nobel Prize for explaining why credentials carry information at all.

His 1973 model starts from a problem employers still have: you cannot observe how capable someone is before hiring them. His answer was that candidates invest in things that are visible and expensive, and the expense is what makes them informative. A degree does not necessarily make anyone more productive. It works as a signal because completing one is harder for a less capable person than for a more capable one. The difference in cost is the message.

A tailored application worked the same way in miniature. Reading the posting properly, thinking about whether you fit, and writing an argument for it took a couple of hours. Those hours cost more if you did not have much to point at.

What the evidence shows changed

Anaïs Galdin at Dartmouth and Jesse Silbert at Princeton tested this directly. In Making Talk Cheap: Generative AI and Labor Market Signaling, they use around 61,000 postings, 2.7 million applications, and 212,000 workers from Freelancer.com between January 2021 and July 2024.

Anaïs Galdin at Dartmouth and Jesse Silbert at Princeton tested this directly on Hire Freelancers & Find Freelance Jobs Online , using around 61,000 postings, 2.7 million applications and 212,000 workers between January 2021 and July 2024.

They scored applications on how much genuine effort each one revealed. Before large language models, the average score was 5.16. For AI-written applications afterward, it was 11.88. Length went from 89 words to 155.

Employers adjusted. Before, an application one standard deviation stronger on that measure improved a worker's hiring odds about as much as bidding $26 less on the job. That relationship weakened sharply afterward. When the authors modeled the market with the signal removed entirely, average wages fell about 5 percent and hiring dropped 1.5 percent per posting. Workers in the top ability quintile were hired 19 percent less often. Those in the bottom quintile were hired 14 percent more often.

The market did not only get noisier. It got less accurate in a specific direction, and the people who lost ground were the ones the old signal used to identify.

The part CVs never handled

There is a second problem, and it is older than any of this.

A CV is a good record of assignments. Titles, dates, employers, formal credentials: these are checkable facts, and they matter.

It is a poor record of how the work went. Whether you saw the flaw in the plan before anyone else. Whether colleagues brought you problems early because you were useful to think with. Whether the project survived because someone renegotiated the scope in week three and that someone was you. Whether people joining your team became productive faster than people joining the team next door.

You can write "strong collaborator with sound judgment" on a résumé. Anyone can write it, which is why it has never carried information.

What still costs something to produce

A useful filter for the current market: where text generation is close to free, the signals that hold their value are the ones whose cost is still someone's time, attention, and reputation.

Colleagues fit that description. A former manager, a peer on the same project, someone who reported to you: each of them spent months watching you work under ordinary conditions, and none of them can delegate that to a model. What they know took real time to acquire, and putting their name behind it, even anonymously, puts something of theirs at stake.

The familiar objection is that this does not scale. Reference calls are slow, awkward, and often produce nothing beyond confirmation of employment dates. That is fair, and it explains why the practice stayed marginal even though its informational value was never in doubt.

PROVEN was built around that constraint. Rather than one phone call to a referee the candidate picked, it gathers structured, anonymous input from people who worked with someone directly, weights responses according to the working relationship, and reports how many people contributed and with what measured reliability. It does not remove bias, and it does not tell anyone who to hire. It adds a source of information that is harder to manufacture on demand than a document.

What the CV is being asked to carry

For its Future of Jobs Report 2025, the World Economic Forum asked employers which skills they consider core today. Analytical thinking came first, with seven in ten calling it essential. Then resilience, flexibility and agility, then leadership and social influence, creative thinking, and motivation and self-awareness. Empathy and active listening made the top ten, alongside curiosity and lifelong learning. Employers expect 39 percent of core skills to change by 2030.

Now look at a CV template. It has a field for your title and one for your degree. Nothing on it records whether you hold steady under pressure, whether people follow you when you have no authority over them, or whether you keep learning once nobody is grading you.

All of that was visible, repeatedly, to the people working next to you. The information already exists. It has simply never been collected.

The capabilities employers say they need most are the ones professionals currently have the fewest ways to demonstrate. Closing that gap would do more for careers over the next decade than any further improvement in how documents get written.

PROVEN validation is free for professionals. If you want to see what colleagues would confirm about your work, you can start one in about five minutes