---
id: "lib-2026-09-24-the-evaluation-bottleneck-in-hiring"
type: "library"
title: "The Evaluation Bottleneck in Hiring"
lang: "EN"
org: "arXiv"
source_type: "Paper"
date: "2026-09-24"
week: "2026-W39"
url: "https://nosov.ai/library/lib-2026-09-24-the-evaluation-bottleneck-in-hiring/"
original_url: "https://arxiv.org/pdf/2609.30058"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "A formal economic model from Stanford and Cornell researchers including Jon Kleinberg, built on recent empirical findings, but a preliminary theory draft with no new data."
summary_basis: "full-text"
original_minutes: "90"
tags: ["hiring", "entry level"]
---

# The Evaluation Bottleneck in Hiring

Tier A (Strong evidence): A formal economic model from Stanford and Cornell researchers including Jon Kleinberg, built on recent empirical findings, but a preliminary theory draft with no new data.

Summary based on the full text.

## In short

A September 2026 arXiv working paper (preliminary draft) by Itai Ashlagi and Ramesh Johari (Stanford), Jon Kleinberg (Cornell) and Anushka Murthy. It models what happens to hiring when AI makes CVs and cover letters cheap to produce and less informative. The firm rationally leans on visible experience, which hurts good candidates without a track record. A cheap intermediate assessment before full screening can restore their chances.

## Key numbers

- 4: applicant types modelled (experience x fit); inexperienced-but-compatible applicants lose most as AI saturation rises
- 2: market failures identified: the firm screens no one, or screens only experienced applicants

## What it reports

- As application materials become noisier, a Bayesian firm puts less weight on them and more on prior experience; the amount of evidence an inexperienced applicant needs to beat an experienced one rises with AI saturation.
- Of four applicant types (experienced or not, compatible or not), inexperienced-compatible applicants suffer the largest fall in average score; incompatible applicants of both groups gain.
- The firm screens only while expected value exceeds screening cost and ignores the worker's gain from a match, so it under-screens relative to the social optimum.
- Two failures follow at high AI saturation: no applicant is screened even though a planner would screen one, or only experienced applicants are screened.
- Adding a cheaper intermediate step (work sample, live skill test, structured first interview) is never worse for the firm and is strictly better when some applicants sit between "worth testing" and "worth full screening".
- At high AI saturation, nearly all submitted inexperienced-compatible applicants gain a positive-probability path to screening under multistage hiring.
- The paper cites empirical work showing AI-tailored applications raised callbacks but lost predictive value, pushing employers toward work history.

## How they got it

A game-theoretic model: one firm, one vacancy, applicants with observable experience and hidden match quality. AI saturation lowers application cost, raises the noise in application materials and raises effective screening cost. The firm ranks applicants by posterior fit and screens sequentially (a version of Weitzman's Pandora problem); results are proven as propositions.

## Limits

- Pure theory: no data, no calibration, no estimated magnitudes.
- Static, single-vacancy setting; long-run effects on who accumulates experience are discussed but not modelled.
- Assumes AI degrades signals uniformly; real effects may differ by role and tool.

## Why it matters for companies

If AI-written applications all look alike, screening by CV drifts toward "has done this job before", and companies lose access to capable early-career people. The paper gives a design answer that HR can act on: a short, verifiable task early in the process, cheaper than a full interview loop. That step also gives the company evidence that AI-polished text cannot provide. Firms worried about pipeline for junior roles can treat assessment design, not application volume, as the bottleneck.

Original: https://arxiv.org/pdf/2609.30058