Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2607.28222 Date read: 2026-09-02 Connected to: L-004, L-008 Kind: content Escalation: store-only Escalation rationale:

What this is

A large-scale field experiment (N=70,000) comparing human vs. AI voice-agent job interviews, with hiring decisions made by humans post-interview. The paper reports that AI-interviewed applicants receive 12% more job offers and show higher retention, attributed to reduced variance in interview conditions.

What I took from it

The work is primarily an empirical validation that automation reduces variance in information collection, producing measurable downstream benefits (job offer rates, retention). This is straightforward optimization through standardization—not a mechanism inquiry.

However, the structure does illustrate L-004 and L-008 in a clean way: the AI agent creates a legible proxy (standardized interview behavior, tone, question sequencing) that human recruiters then optimize around. The 12% uplift likely reflects that recruiters, evaluating the interview record rather than the raw applicant, are making decisions based on a more consistent signal. The paper does not investigate whether this consistency selects for actual job performance or merely interview-game fitness—a classic Goodhart candidate. The lack of decline in retention suggests the proxy may be reasonably aligned, but this is not proven.

The generalizability claim (variance reduction → better outcomes) is domain-specific to hiring, and the paper offers no mechanism that would hold across protocol systems broadly.

Research connections

L-004: Confirms Goodhart risk: the consistency proxy created by AI agents becomes a legible optimization target for recruiters; the paper does not measure proxy divergence from true job fitness.
L-008: Illustrates the setup: computable enforcement (standardized interview) creates a legible signal that downstream decision-makers optimize around; no evidence of downstream gaming or perverse adaptation yet.
seed-062: Light touch: formalization of interview (AI standardization) increases legibility; the paper does not examine whether this changes what recruiters attend to.

Seed

Seed title: none

Seed type: —

Seed text: —