Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Understanding Design Fixation in Generative AI

Liuqing Chen, Yaxuan Song, Chunyuan Zheng, Qianzhi Jing, Preben Hansen, Lingyun Sun

TL;DR

This paper addresses the emergence of GenAI design fixation, the tendency of generative models to bias toward familiar design spaces and reduce novelty. It proposes a theoretical framework defining GenAI design fixation, identifying causes (data bias, architecture) and human factors, and detailing manifestations in text and image generation. An empirical study with ten novice designers using GPT-4o and Midjourney demonstrates concrete fixation patterns in both modalities and their impact on human-AI co-ideation, supplemented by comparisons to a Red Dot design dataset. The authors propose multi-faceted mitigation strategies for creativity support tools, including diversifying sources, improved prompt engineering, interactive interfaces, education, and comprehensive evaluation metrics, aiming to preserve designer agency and enhance creative outcomes.

Abstract

Generative AI (GenAI) provides new opportunities for creativity support, but the phenomenon of GenAI design fixation remains underexplored. While human design fixation typically constrains ideas to familiar or existing solutions, our findings reveal that GenAI similarly experience design fixation, limiting its ability to generate novel and diverse design outcomes. To advance understanding of GenAI design fixation, we propose a theoretical framework includes the definition, causes, manifestations, and impacts of GenAI design fixation for creative design. We also conducted an experimental study to investigate the characteristics of GenAI design fixation in practice. We summarize how GenAI design fixation manifests in text generation model and image generation model respectively. Furthermore, we propose methods for mitigating GenAI design fixation for future creativity support tool design. We recommend adopting the lens of GenAI design fixation for creativity-oriented HCI research, as the unique perspectives and insights it provides.

Understanding Design Fixation in Generative AI

TL;DR

This paper addresses the emergence of GenAI design fixation, the tendency of generative models to bias toward familiar design spaces and reduce novelty. It proposes a theoretical framework defining GenAI design fixation, identifying causes (data bias, architecture) and human factors, and detailing manifestations in text and image generation. An empirical study with ten novice designers using GPT-4o and Midjourney demonstrates concrete fixation patterns in both modalities and their impact on human-AI co-ideation, supplemented by comparisons to a Red Dot design dataset. The authors propose multi-faceted mitigation strategies for creativity support tools, including diversifying sources, improved prompt engineering, interactive interfaces, education, and comprehensive evaluation metrics, aiming to preserve designer agency and enhance creative outcomes.

Abstract

Generative AI (GenAI) provides new opportunities for creativity support, but the phenomenon of GenAI design fixation remains underexplored. While human design fixation typically constrains ideas to familiar or existing solutions, our findings reveal that GenAI similarly experience design fixation, limiting its ability to generate novel and diverse design outcomes. To advance understanding of GenAI design fixation, we propose a theoretical framework includes the definition, causes, manifestations, and impacts of GenAI design fixation for creative design. We also conducted an experimental study to investigate the characteristics of GenAI design fixation in practice. We summarize how GenAI design fixation manifests in text generation model and image generation model respectively. Furthermore, we propose methods for mitigating GenAI design fixation for future creativity support tool design. We recommend adopting the lens of GenAI design fixation for creativity-oriented HCI research, as the unique perspectives and insights it provides.

Paper Structure

This paper contains 38 sections, 2 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Conceptualizing GenAI design fixation
  3. Preliminary causes analysis
  4. Distinctions and correlations between GenAI design fixation and human design fixation
  5. Understanding GenAI design fixation in practice
  6. The aim of this study
  7. Participants
  8. Procedure
  9. Data collection and analysis
  10. Data collection
  11. Data analysis
  12. Text generation analysis
  13. Image generation analysis
  14. Findings
  15. Manifestations in text generation design fixation
  16. ...and 23 more sections

Figures (5)

  • Figure 1: The process and dynamics of fixation within GenAI systems and how it correlates with human fixation in design processes.
  • Figure 2: The process of participants engaging in our experiment.
  • Figure 3: Comparison of the top 10 most frequent word stems in design solutions generated by ChatGPT and those recognized in Red Dot award-winning designs. Common terms across both datasets are highlighted and include "comfort", "light", "ergonomic", "innovative", and "aesthetic".
  • Figure 4: Visualization of t-SNE dimensionality reduction applied to the embeddings from Midjourney-generated chair images.
  • Figure 5: Manifestations of design fixation on image generation models from our experiment displayed on the left.