Innovation & Tech

AI Has Broken the Link Between Output and Progress

AI has cut the cost of producing applications, papers and code while review capacity stays human-paced. Rising output no longer signals progress.

By Olivia Hart

4 min read

More output is no longer a sign of progress
More output is no longer a sign of progressStartupStockPhotos / pixabay

What's News

  • Average job ads on a leading platform draw ~254 applicants; recruiter application loads are up 412%.
  • arXiv received 40,363 submissions in September — a record, double the figure from two years earlier.
  • A study across 24 occupations found candidates using the employer's own AI model were 23–60% more likely to be shortlisted.
  • OpenAI published AI-generated solutions to 370+ outstanding math problems in October.
  • Greenhouse CEO Daniel Chait says candidates using its one-per-month priority flag are hired at ~5x the rate of others.

By Olivia Hart

The average job posting on one leading hiring platform now draws around 254 applicants, and the workload per recruiter is up 412% — while the number of people available to judge those applications has barely moved. That asymmetry, argues a new analysis, is quietly breaking the cue most organizations use to measure progress: how much they produce.

The piece, built around the biology concept of an "evolutionary trap," makes a blunt claim. AI has changed the environment faster than the systems operating in it — hiring, publishing, government — can adapt.

What is an evolutionary trap?

A newborn sea turtle follows one rule: crawl toward the brightest horizon. For most of turtle history, that was the sky over the sea. Then humans built streetlights and hotels. This year on Florida's Anna Maria Island, volunteers recorded more than 3,000 hatchlings heading the wrong way; 55 were found dead on a road in August.

The business analogy is pointed. Systems are being used exactly as intended. The environment around them has shifted, and the signals they steer by — above all, rising output — no longer point anywhere useful.

Where is the volume crushing institutions?

The evidence spans three sectors:

  • Hiring: 254 applicants per ad on average, with recruiter application loads up 412%, per the hiring platform Greenhouse.
  • Government: In May, the U.K.'s Information Commissioner's Office issued new guidance to help public authorities cope with AI-generated freedom of information requests.
  • Research: In September, the preprint server arXiv received 40,363 submissions — a record, double the figure from two years earlier. arXiv itself noted there used to be "an easily discernible, practical limit" to how fast submissions could be produced.

AI destroyed that limit by cutting the marginal cost of production and lowering the knowledge barrier for new producers. The catch: everything produced still needs review, and review cannot scale the way production can. Producing something requires only that it seem right; reviewing it means deciding it is right. Final judgment stays with humans.

That is the core asymmetry, and the article does not pretend AI can fix it on the review side.

What happens to signals of quality?

Nobel laureate Michael Spence, who shared the 2001 economics prize for signaling theory, showed that signals carry information when they are expensive to send. A bespoke cover letter once cost real effort, so it meant something.

AI removes the cost. The problem is not that AI-assisted applications are bad — it is that they are good and uniformly polished, so a serious candidate becomes hard to distinguish from someone who auto-applied to a thousand jobs overnight.

Outsourcing the sorting to AI is no clean fix:

  • One study simulating hiring across 24 occupations found candidates who used the same AI model as the employer were 23% to 60% more likely to be shortlisted than equally qualified people who wrote their own résumés.
  • Researchers found three state-of-the-art AI models favored résumés with white-associated names 85% of the time.
  • A federal court has allowed a discrimination lawsuit over Workday's AI screening tools to proceed.

AI screeners reward the polish that no longer means anything, and they inherit old prejudices. That is not a sorting solution; it is a sorting problem wearing one.

Does more output mean more understanding?

In October, OpenAI published AI-generated solutions to more than 370 outstanding math problems. Mathematician Daniel Litt's caveat deserves the last word on the hype: "to extract understanding from [the OpenAI results] there will be a huge amount of human labor involved."

Software engineers have a name for the gap: comprehension debt — the distance between how much code exists and how much anyone actually understands. Sometimes the gap is harmless. An accurate forecast is useful regardless of how it was made. But when conditions change, a business can keep acting on a model's confident answers long after the assumptions behind them have expired.

Accountability compounds this. A company remains answerable for decisions it hands to machines, and a regulator or court will expect an explanation. Terence Tao adds a subtler cost: often the process of reaching a solution, not the solution itself, advances understanding. Skip the working-out, and you may skip your next breakthrough.

What is scarce now?

The article's answer: not production capacity, but review capacity. Every AI-assisted memo or slide deck creates work for whoever must judge whether it is useful.

One design fix is already live. Greenhouse lets job seekers flag one application per month as top priority. The flag works because it costs the sender something AI cannot reduce — the month's only flag. CEO Daniel Chait says candidates who use it are hired at roughly five times the rate of everyone else.

The same logic applies internally: when effort no longer caps output, leaders must design limits in, forcing people to choose what most deserves attention.

The caveat the piece leaves only partly explored: judgment itself depends on people experienced enough to tell good from merely plausible. An organization that hands its reasoning to AI models will, over time, have fewer people able to judge what the machines produce. The hatchlings headed for the streetlights; the warning for managers is that output growth is now the brightest — and least reliable — horizon on the beach.

Original: mysuncoast.com

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,205 articles

Related articles

« Previous articleNext article »