Homogenization Effects of Large Language Models on Human Creative Ideation

TL;DR

The paper tests whether LLM-based CSTs homogenize human creative ideation, comparing ChatGPT to Oblique Strategies in a within-subject design across divergent tasks. It finds significant group-level homogenization with ChatGPT, alongside increased idea generation and improved fluency, flexibility, and elaboration, but no clear gain in originality and a reduced sense of personal ownership over ideas. The authors argue that homogenization arises from low inferential distance and shared model outputs, not from individual fixation, and propose design and algorithmic strategies to mitigate it, such as output obliqueness and diverse decoding. The work contributes a quantitative homogenization framework, demonstrates cross-domain effects beyond writing, and guides CST designers toward intent elicitation and tool diversity to preserve originality. The findings have practical implications for deploying LLM-based CSTs in creative workflows and for future research on reducing output conformity in AI-assisted ideation.

Abstract

Large language models (LLMs) are now being used in a wide variety of contexts, including as creativity support tools (CSTs) intended to help their users come up with new ideas. But do LLMs actually support user creativity? We hypothesized that the use of an LLM as a CST might make the LLM's users feel more creative, and even broaden the range of ideas suggested by each individual user, but also homogenize the ideas suggested by different users. We conducted a 36-participant comparative user study and found, in accordance with the homogenization hypothesis, that different users tended to produce less semantically distinct ideas with ChatGPT than with an alternative CST. Additionally, ChatGPT users generated a greater number of more detailed ideas, but felt less responsible for the ideas they generated. We discuss potential implications of these findings for users, designers, and developers of LLM-based CSTs.

Figures (6)

• Figure 1: Participant responses were more homogenous at the group level (i.e., more semantically similar to the average embedding of all participant ideas) when using ChatGPT.
• Figure 2: Participant responses were not observably more or less homogenous at the individual level (i.e., more semantically similar to the average embedding of this participant's ideas) when using ChatGPT.
• Figure 3: Participants assigned less creative responsibility to themselves, and more to the CST, when working with ChatGPT.
• Figure 4: Participants generated more ideas with ChatGPT than with OS.
• Figure 5: Participants generated ideas across more categories with ChatGPT than with OS.