ORC IAP Seminar 2026 Talk 6 Gili Rusak
Gili Rusak
PhD Student
Harvard University
Title
A Paragraph is Worth 1,000 Choices: Towards Scalable Preference Elicitation with Artificial Intelligence
Abstract
We elicit preferences from workers in an online labor market over varying numbers of real paid jobs. When workers evaluate only a few jobs, their reported rankings closely track underlying preferences. But as the number grows, self-report fidelity deteriorates sharply. We show that AI offers a scalable alternative. Workers write short descriptions of their “tastes”—the kinds of tasks they generally like and dislike. We feed these descriptions to a language model and prompt it to rank jobs on the worker’s behalf. Unlike self-reports, AI-inferred preferences scale: predictive accuracy holds from the 10th job to the 50th and beyond. The AI even generalizes to novel jobs workers never saw. In simulated allocation mechanisms, participants prefer assignments from AI-inferred preferences to those from self-reports. These findings suggest that AI can, at negligible cost, proxy preferences that are practically impossible to elicit—enabling otherwise infeasible market designs.
Bio
Gili is a 4th year computer science PhD student co-advised by Prof. David Parkes, Harvard, and Prof. John Horton, MIT. Gili is interested in questions surrounding how modern AI can be leveraged to enable practical implementation of search and matching mechanisms. Her work is supported by Upwork.