Large language models are homogeneously creative

Area of Science:

• Artificial Intelligence
• Cognitive Psychology
• Human-Computer Interaction

Background:

• Large language models (LLMs) are increasingly marketed as creativity support tools.
• Existing research suggests LLM use can narrow creative output diversity.
• Previous studies focused on single LLM interactions, leaving the generalizability of these findings unclear.

Purpose of the Study:

• To investigate whether the narrowing effect on creativity is specific to individual LLMs or a general characteristic of LLM use.
• To compare the diversity of creative responses generated by humans versus multiple LLMs.

Main Methods:

• Standardized creativity tasks were administered to both human participants and various LLMs.
• Population-level response diversity was statistically analyzed and compared between human and LLM groups.
• Confounding variables were controlled for in the analysis.

Main Results:

• LLM-generated responses exhibited significantly higher similarity to each other compared to human-generated responses.
• This pattern held true across different LLMs and creativity tasks.
• Humans demonstrated greater diversity in their creative outputs than LLMs.

Conclusions:

• The tendency for LLM outputs to converge suggests a general limitation, not specific to any single model.
• Using current LLMs as creative partners may lead users towards homogenous creative outcomes.
• This finding has significant implications for the future development and application of AI in creative fields.