A poetry study found co-creating beats editing AI drafts

The system

The system under review is not a product but a controlled experiment: two studies testing how the design of a generative AI writing interface changes human creativity. Jack McGuire, David De Cremer and Tim Van de Cruys, based at Northeastern University, the National University of Singapore and KU Leuven, built a GPT-4-based poetry system and published the results in Scientific Reports in August 2024, available via the publisher's article page and the PubMed Central full text. The task was poetry writing, not music composition, a distinction worth holding onto before extending the findings anywhere else.

What the documents establish

Study 1 assigned 146 participants to write alone, to edit a GPT-4 poem with sophisticated editing tools, or to have GPT-4 generate a poem outright. Solo writers were rated more creative, by themselves and by external evaluators, than editors of an AI draft. Study 2 added a redesigned 'co-creator' condition, where people generated lines alongside the system rather than revising its finished output. The creativity gap seen for editors disappeared for co-creators. The PMC full text reports that creative self-efficacy, measured on a three-item scale, differed significantly across conditions (F(2,149)=7.01, p=0.001), and mediation analysis identified self-efficacy as the mechanism: co-creators reported self-efficacy statistically indistinguishable from solo writers, while editors reported less.

Craft and rights

The rights question this raises is not about training data but about interface design as a decision made on a musician's behalf. Many commercial generation tools default a user into an editor's role: accept, regenerate, tweak a finished stem or take. This study's own framing suggests that role placement, not just model quality, shapes whether a person experiences a session as their own creative work or as clean-up on someone else's draft. That distinction matters for how a producer credits and values a session, and for whether a tool's design supports or quietly displaces a musician's sense of authorship. This is an editorial extension of the paper's own findings, not a claim the authors make about music specifically.

Outcomes and open questions

The authors themselves flag that their poetry task and Prolific-recruited sample may not generalize, and they call for replication with newer systems such as GPT-4o or Claude 3.5. Nothing in the paper tests musical tasks, professional creators, or long-term use. Whether an interface framed as co-creation in one domain produces the same self-efficacy effect in audio tools, where 'editing' a generated stem is often the entire workflow, remains untested.

• Does this tool put me in the role of generating alongside the system, or only revising what it already finished?
• Would I rate this session as more mine if the interface asked for my input earlier in the process?
• What population and task was actually studied, and how far does that stretch toward my own instrument or genre?

The result is narrow but concrete: in this poetry task, a co-creator interface preserved a sense of creative capability that an editor interface did not, and that difference tracked measured self-efficacy rather than the model's raw output quality.