Research

Working Papers

  • Learning to Prompt: Human Adaptation in Production with Generative AI (JMP)
    Presentations: University of Toronto 2025
    What is the role of human input in AI-assisted production? Humans interact with generative AI through combinations of words called prompts. A key feature of human input is adaptation: users dynamically modify their prompts based on their understanding of AI. I empirically investigate two types of adaptation: (1) adaptation to new AI versions, referring to how people change their prompts in response to AI upgrades; (2) adaptation to outputs from previous prompts, referring to how people adjust their prompts iteratively to converge on desired outcomes. I study this adaptation using prompt-level data from Midjourney, a leading AI image generator. First, users adapt to AI upgrades by writing different words in their prompts. By submitting prompts written for the old version to the new AI and vice versa, I decompose the output shifts as arising from prompt changes (73%), AI changes (20%), and an AI-human cross effect (7%), implying complementarity between AI and human inputs. Second, prompts evolve within the creative process of an artwork. I estimate a structural model of the creative process using the sequential search framework. Counterfactual shows that without human adaptation, users need three times more prompts to achieve data-observed results. Both results highlight the importance of human judgment and adaptation in the creative process.

  • Hiding From Generative AI
    Revise and Resubmit, Quantitative Economics
    Presentations: TSE Digital Economics Conference 2025; University of Toronto 2024; NASM-ES 2024; ESIF Economics and AI+ML 2024; EEA-ESEM 2024; ALEA 2024; CEA 2024
    How does generative Artificial Intelligence (AI) change the behaviors of content creators? I investigate the effect of an AI image generator on artists’ incentives to publish artworks using data from an online art platform, DeviantArt. On November 11 2022, DeviantArt introduced a generative AI image generator into the platform and artworks on this platform entered training data by default. Using a difference-in-differences estimation with artists who do not use AI, I show that digital artists publish 21% fewer artworks following AI’s introduction on this platform, in contrast to artisan crafts artists. This reduction could potentially hinder knowledge spillovers to other artists and AI training data availability. By matching the artworks of artists who publish both on DeviantArt and Instagram, I find that despite artists publishing fewer artworks on DeviantArt, the quality of published artworks for a given artist remains the same after the introduction of AI.

Work In Progress

  • Knowledge Spillovers in the Diffusion of Generative AI
  • Do Human Users Correct AI-Created Stereotypes?
    with Ruiqi Sun (University of Hong Kong) and Siyuan Liu (University of Toronto)
  • Platforms in Platform
    Show Abstract

    Some apps on platforms are themselves smaller platforms featuring their own app stores. All else equal, buyers are better off on a platform with more sellers and sellers are better off on a platform with more buyers. However, different buyers and sellers may prefer to be active on different platforms. Buyers may choose to use different platforms to access more targeted products, while sellers may use different platforms to avoid competition for buyers' attention. This paper argues that segmenting the market using smaller platforms can be profitable for the large platform, as it can charge different platform fees to sellers of different quality and keep buyers' attention to better sellers. This paper also shows conditions where the large platform might ban the smaller platform even though its presence would be socially beneficial.