Does AI assistance help individual writers or hurt collective creativity?

In a preregistered two-phase online experiment, 293 participants wrote an eight-sentence short story for a "teenage and young adult audience," randomly assigned to write with no AI assistance, with one GenAI-generated idea, or with five GenAI-generated ideas. Evaluators then rated each story on novelty and usefulness. The paper reports that "access to generative AI ideas causes stories to be evaluated as more creative, better written, and more enjoyable, especially among less creative writers." At the same time, "generative AI–enabled stories are more similar to each other than stories by humans alone." The authors summarize the combination as "an increase in individual creativity at the risk of losing collective novelty."

The paper frames two competing pathways by which GenAI ideas could affect writing, and finds evidence consistent with both operating together. GenAI ideas can work as a "springboard," giving a writer "potential starting points that can result in a 'tree structure' of different storylines" and helping overcome "writer's block and the fear of a blank page" — the individual-level gain. But the same starting point can anchor the writer, since "the output offered by generative AI may be derivative and thus not provide a fertile ground for new and creative ideas," so many writers converge on similar material — the collective-level cost. The authors name the resulting structure directly: "this dynamic resembles a social dilemma: With generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced."

This runs parallel to Does AI assistance homogenize or preserve creative diversity?, which reports the same individual-versus-collective split — AI ideation raising individual ratings while shrinking the pool — in a different task (creative metaphors) and a different manipulation (ideation versus refinement rather than zero, one, or five seed ideas). Two independent experimental designs landing on the same social-dilemma structure, across short fiction and metaphor writing, is evidence the pattern is not an artifact of either task. A similar concentration-under-AI-assistance shape shows up outside human writing too: Do AI research agents explore as broadly as human researchers? finds agent-generated research ideas cluster more tightly than human papers, and Do AI writing assistants push non-Western writers toward Western styles? finds a related homogenizing pull toward a dominant style rather than toward a shared idea pool. Can statistical rarity measure whether stories are truly original? offers a measurement route — statistical rarity in a feature space of narrative decisions — that this paper's "more similar to each other" finding does not itself quantify.

The excerpt does not report the effect size or statistical test behind "more similar to each other," how similarity was operationalized, or whether the effect scales with one versus five GenAI ideas. It does not establish that anchoring, rather than some other property of LLM-generated ideas, is the actual mechanism, since the paper offers anchoring and derivativeness as explanations rather than as tested causes. The design is a single short-fiction task, one language, and one age-targeted audience, so extending the result to other creative domains is extrapolation, not finding. Within those limits, the paper is reasonably read as evidence that a GenAI-ideation intervention which looks beneficial from inside any one writer's results can still shrink what the population of writers produces as a whole.