LLMs Homogenize Creative Ideas, Study Shows
NeurIPS 2022 study finds ChatGPT users generate more similar ideas on creative tasks than others, with greater detail but less ownership—risking 'algorithmic monoculture' from shared models.
LLM-Assisted Creativity Produces Homogeneous Outputs
A NeurIPS 2022 paper demonstrates that when humans use ChatGPT for creative ideation, their outputs converge toward similarity, unlike ideas from users of other tools. Participants using LLMs produced more detailed concepts but felt less responsible for them, reducing psychological ownership. This happens because everyone draws from the same training data and algorithmic patterns in centralized models like ChatGPT or Claude, creating an 'algorithmic monoculture' where diverse, quirky human ideas get averaged out into predictable variations.
The author confesses to relying on ChatGPT (nicknamed 'Chad') for naming his newsletter 'TechTonic Shifts,' highlighting how common this pitfall is—even for experienced users expecting wild innovation from group LLM brainstorming.
Trade-offs: Sparks vs. Full Ideas
LLMs excel at kickstarting detailed thoughts but undermine true creativity by funneling users into shared idea spaces. The result: a bland creative landscape lacking unexpected edges that define human originality. To counter this, redesign AI tools as a 'creative compass'—providing initial sparks or prompts rather than complete ideas. This preserves user agency, fosters ownership, and encourages embracing personal quirks over algorithmic averages.
In practice, treat LLMs as idea igniters for production workflows, but rely on human imagination for differentiation when building AI-powered products or content.