Research Interest
Research: Quantitative Marketing, Generative Models, Creativity, Aesthetic Design
Methods: Field Experimentation, Deep Learning
Research Interest
Research: Quantitative Marketing, Generative Models, Creativity, Aesthetic Design
Methods: Field Experimentation, Deep Learning
Working Paper
Guided Creativity: AI Intermediation for Enhancing Originality and Quality in Visual Design
Joint work with Artem Timoshenko and Guy Aridor
Creative fixation is a common side effect when designers seek inspiration from successful designs, often limiting the originality of subsequent work. This paper introduces AI intermediation, a novel approach that leverages generative models to overcome this challenge. Our approach creates variations of leading designs that maintain core semantic concepts while differing visually, and provides these variations to the designers instead of the original exemplars. This allows communicating valuable insights and inspiring novel interpretations without inducing fixation. We empirically validate the proposed approach using a field experiment involving professional designers in a logo design contest. Results show that designers with AI intermediation produce (1) higher quality work than those with no exposure to exemplars, and (2) more-original work than those with direct exposure to exemplars. We further decompose the sources of creativity and demonstrate that while the generative model yields distinct variations, human creativity remains pivotal for improving originality and quality of the final designs. Consequently, AI intermediation presents an efficient facilitator to human-driven creative process.Â
Work in Progress
Using Context-Aware Generative AI to Close Cross-Cultural Design Gap
When serving brands in foreign markets, designers face a cultural gap that challenges their understanding of market conventions and thus the creation of effective design. Using data from a logo design contest platform, we show that the cultural gap has a negative effect on design ratings. One mechanism we document is that foreign designers fail to balance the familiarity and novelty when combining design elements. Motivated by these findings, we develop a context-aware generative pipeline that constructs inspiration sets reflecting the target market's conventions. The pipeline extracts the co-occurrence structure of design elements from real logos of the market, samples element combinations that reproduce this structure, and renders them into images through a fine-tuned text-to-image model. We conduct an experiment to evaluate whether context-aware inspiration improves the quality of designs produced by foreign designers.