Shared AI prompts lifted stories but narrowed them together

The process wall

A writer facing a blank page sometimes uses a generated line for inspiration rather than a finished draft. A study published in Science Advances on 12 July 2024, by Anil R. Doshi and Oliver P. Hauser, isolates exactly that decision, holding a story's length, audience and topic constant so the only variable is how much AI-generated inspiration, if any, a writer drew on.

What the documents show

In the first phase, 293 writers each produced an eight-sentence story ‘appropriate for a teenage and young adult audience,’ randomly assigned to a Human-only condition, a Human-with-one-idea condition, or a Human-with-five-ideas condition, where GPT-4 supplied only ‘a three-sentence starting idea’ to inspire the writer's own story. In the second phase, 600 separate evaluators each read six randomly assigned stories, without being told which condition produced them, rating novelty, usefulness and emotional qualities before being told whether AI was involved. The paper reports that access to generative ideas made stories ‘more creative, better written, and more enjoyable, especially among less creative writers,’ but that ‘generative AI-enabled stories are more similar to each other than stories by humans alone’ — the authors describe this pairing as resembling ‘a social dilemma.’

Authorship and limitation

The mechanism to name precisely is about distribution, not capability: this is not evidence that a generative tool cannot support creative work, it is evidence about what happens to the range of outputs when many writers draw on the same generator for the same narrow prompt. A studio issuing one tool to an entire writers' room, on this design, risks the convergence the study measured even as each individual writer's own story reads as more polished. The authors' intention-to-treat analysis — comparing by assigned condition regardless of whether a writer actually used the offered idea — is named in the paper specifically to support treating the effect as causal rather than incidental.

What survives

The authors' own limitations state the task was ‘constrained in its length … medium … and type of output,’ with ‘no interactiveness with the LLM or variation in prompts,’ so the result may not generalise to a longer form, a different medium, or freer prompting. Editorially, a studio wanting individual quality without collective sameness might vary which tool, which prompt, or how many options each writer receives, though the paper itself did not test whether that would work.

• Check whether a writers' room or contest gives every participant the identical generative prompt, since that is the condition the study found narrows collective output.
• Weigh the study's own limits on task length and medium before applying its finding to a longer or differently-mediated project.
• Have evaluators rate submissions before learning which used AI assistance, mirroring the study's blind-then-revealed design.

Individual gain and collective narrowing appear together in this design, and the authors' own account treats that pairing, not either half alone, as the finding a studio should plan around.