Making Talk Cheap – Generative AI and Labor Market Signaling [pdf] (jesse-silbert.github.io)

Researchers using Freelancer.com data show that large language models (LLMs) have eroded the signaling value of customized written applications and materially altered hiring equilibria. Galdin and Silbert build an LLM-based measure of how tailored a proposal is to a job posting, combine it with click/time data (including platform adoption of an AI proposal tool in April 2023), and document that before LLMs employers paid a premium for customized proposals (a one‑SD higher signal equaled a $26 bid advantage) and signals predicted worker effort and success. After mass LLM adoption, employers’ willingness to pay for customization collapses, AI‑written proposals decouple effort from signal, and signals no longer predict successful completion. To quantify equilibrium effects from signaling loss alone, the authors estimate a structural model that fuses a Spence‑style signaling framework with a discrete‑choice employer demand model and a multi‑dimensional scoring‑auction view of applications. Identification exploits equilibrium hiring probabilities conditional on bid and effort to recover worker beliefs, costs, and abilities. A counterfactual where writing costs fall to zero shows a substantial decline in meritocratic sorting: top‑quintile workers are hired 19% less often while bottom‑quintile hires rise 14%. The paper signals important implications for platform design, hiring practices, and the need for new verifiable signals if generative AI continues to make “talk cheap.”