Does ChatGPT make different users converge on similar ideas?

A 36-participant within-subjects study compared ChatGPT to the Oblique Strategies (OS) card deck as creativity support tools across four divergent-ideation prompts drawn from the Torrance Tests of Creative Thinking. Participants produced 1,271 ideas total, split between the two tools. The study's first finding (RQ1) was that "different users tended to produce less semantically distinct ideas with ChatGPT than with" the OS deck — homogenization at the group level. Its second finding (RQ2) was "no difference": the set of ideas any one participant generated was about as diverse in the ChatGPT condition as in the OS condition. The quantity increase from ChatGPT, the paper notes, "did not result in a commensurate increase in diversity as might be expected if diversity is a linear function of quantity."

The authors locate the homogenizing pressure at the group level, not the individual level: "homogenization stems not from individual-level increases in fixation when working with the LLM, but from group-level suggestion of similar ideas to different users by the LLM." They rule out creative fixation — users "becoming fixated on specific subsets of the space of all possible ideas" — because even participants who called ChatGPT a "very fast and accurate" authority still proposed ideas unlike its suggestions, sometimes noting the model "went really specific really quickly" or became "repetitive." Instead they point to "low inferential distance between LLM outputs and apparently finished creative products": because ChatGPT's suggestions already look complete, different users converge on the same finished-looking answers rather than treating them as raw material to diverge from.

This sharpens Does AI assistance homogenize or preserve creative diversity? by locating exactly where AI-ideation's pool-compression happens: at the level of what different users are shown, not at the level of what any one user can think of. It matches the same individual-quality-vs-collective-sameness dissociation found in Can LLMs generate more novel ideas than human experts? and in Why do LLMs generate novel ideas from narrow ranges?, giving that pattern a controlled, non-AI-tool baseline rather than an aggregate statistic alone. It is a different mechanism from Do different AI models actually produce diverse outputs?, which homogenizes outputs across different models; here a single model homogenizes outputs across different human users of that one model.

The study ran in a lab, with short fixed response windows and ideation prompts the researchers supplied rather than ones participants chose themselves, so it does not establish how homogenization plays out in open-ended, self-directed use. It used one general-purpose chatbot (ChatGPT 3.5) rather than a CST built for divergent ideation, and the paper notes prior work finding that "feedback-tuned LLMs may induce creative homogenization where base (non-feedback-tuned) LLMs do not" — meaning this result characterizes RLHF-style chat models specifically, not LLMs as such, and leaves open whether differently tuned or differently designed tools would homogenize less.