One writing assistant nudged many writers toward similar ideas

The writer's problem

A writer who brainstorms with a popular chatbot relies on the same tool countless other writers also turn to, which raises a question no single manuscript can answer: whether the tool is quietly narrowing the range of ideas circulating across many different people's drafts, not just shaping any one person's. A study accepted to the ACM Creativity and Cognition conference by Barrett Anderson, Jash Hemant Shah and Max Kreminski designed an experiment around exactly that possibility.

What the documents show

The paper's own methods describe a within-subjects study in which 33 analyzed participants used two tools for open-ended idea generation: ChatGPT, and the Oblique Strategies deck, a set of prompt cards created by Brian Eno and Peter Schmidt. Using sentence-embedding similarity to measure how close each participant's ideas were to other participants' ideas, the authors report that different users produced less semantically distinct ideas with ChatGPT than with the card deck: ChatGPT users converged on more similar territory. At the same time, ChatGPT users generated more numerous and more detailed ideas, and reported feeling less responsibility for the ideas they produced. A second, separate paper measured a related but distinct effect directly in model output rather than in user behavior. A study of what its authors call generative monoculture compared LLM-generated book reviews against the Goodreads review corpus those models likely trained on, finding that reviews generated by Llama-2-chat clustered in an overwhelmingly positive, far narrower sentiment range than genuine reader reviews, and that simple prompting or sampling changes did not fix the narrowing. One paper shows homogenization across co-writing users; the other shows a generating model's own output can be narrower than the human data it drew from.

The editorial choice

Neither paper tells a writer to stop using a chatbot for brainstorming. Together they suggest an editorial practice: treat a single model's suggestions as one narrow slice of possibility rather than a representative sample of what could be written, and deliberately seek contrasting prompts, sources or tools before settling on a direction, especially on any assignment where distinctiveness from other writers' output using the same tool matters.

What stays with the author

The Anderson study's own finding that ChatGPT users felt less ownership of their generated ideas is a self-report about a lab task, not a rule about any given writer's process. Deciding which idea to pursue, and how much of a machine-suggested option to keep, revise or discard, remains a choice the tools studied here do not make and the papers do not measure as a matter of authorship or credit.

• Are other writers likely to be prompting the same tool toward the same territory?
• Would a second, differently sourced idea-generation step change the direction taken?
• Does felt ownership over an idea track who is credited for the finished piece?

Read together, the two studies describe a mechanism, not a verdict: shared tools can narrow a population's outputs even when they expand any one user's apparent output.