Research
Working Papers
Abstract
(This paper and “Human Adaptation in the Creative Process with Generative AI” were previously circulated together as “Learning to Prompt: Human Adaptation in Production with Generative AI.”)How do humans adapt their behavior in response to AI upgrades, and how does this adaptation affect creative output? I study this question using prompt-level data from Midjourney, a leading AI image generator, where users interact with AI by submitting combinations of words known as prompts. Textual analysis shows that users systematically change how they write prompts following AI upgrades. Consider the image production function as Image = F(AI, Prompt). I ask how much of the change in output is attributable to ∆AI versus ∆Prompt. By submitting prompts written for the old AI to the new AI and vice versa, I decompose the output shifts as arising from prompt changes (73%), AI changes (20%), and a residual (7%), implying complementarity between AI and human inputs. The decomposition should be interpreted as a first-order expansion of the production function. More generally, this paper provides a new way to understand the production function without explicitly assuming a functional form.Human Adaptation in the Creative Process with Generative AI
Abstract
(This paper and “Evolving with Generative AI” were previously circulated together as “Learning to Prompt: Human Adaptation in Production with Generative AI.”)What is the role of human input in AI-assisted production? Humans interact with generative AI through combinations of words called prompts. Users dynamically modify their prompts based on the AI's previous output. I empirically examine how users adjust their prompts iteratively to converge on desired outcomes. I study this creative process using prompt-level data from Midjourney, a leading AI image generator. I document novel facts about prompt patterns in image production and find that they are similar to consumer product searching patterns. (1) Prompts are path-dependent; (2) Prompt length increases step-by-step; (3) Prompts converge to the final prompt along the paths; (4) Users adjust words with greater weights first and then move to less important words. I estimate a structural model of the creative process using the sequential search framework. Counterfactual shows that without human learning in the loop, users need three times more prompts to achieve data-observed results.Abstract
How does generative Artificial Intelligence (AI) impact the incentives of content creators to publish their work? This paper uses a difference-in-differences analysis to identify how much creators withhold their work online to avoid being "stolen" by AI companies for training their models. I investigate the introduction of an AI image generator on an online art platform, DeviantArt. Following this introduction, artworks on this platform entered training data by default. Using an estimation with a majority of incumbent 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 craft artists. This reduction could potentially hinder knowledge spillovers to other artists and AI training data availability. Multi‑homing artists continue to publish at similar levels on Instagram as before, suggesting that they reduce publication specifically on DeviantArt rather than cutting overall production. Furthermore, I find no evidence of a decline in the quality of published works at either the intensive or extensive margins.
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