“It Felt Like Having a Second Mind”: Investigating Human-AI Co-creativity in Prewriting with Large Language Models

“At the intersection of creativity research and HCI, a huge amount of work was dedicated to supporting creativity using computing technologies, notably via the design of Creativity Support Tools (CSTs) [26]. Early work by Shneiderman proposed a four-phase framework to support the development of user interfaces for creative problem solving, which included phases where one Collects ( learn from previous works), Relates (consult with peers and mentors), Creates (explore, compose, discover and evaluate possible solutions), and Donates (disseminate results) [63]. In 2019, Frich et al. conducted a systematic review of previous CSTs developed by researchers since 1999. Through a thematic analysis of these academic papers, they identified 6 stages of creativity that those CSTs aimed to support, including pre-ideation, ideation, evaluation, implementation, iteration and project management [26].”

“Drawing upon previous literature on the creativity [26, 70] and writing [25, 57] processes, we found that LLMs were most intensively used to support three key stages of the entire creativity process: Ideation, Illumination, and Implementation (Figure 1). Participants usually used the LLM for Ideation when they initially had no ideas or only a vague picture in their minds. If they happened to have any thoughts, most preferred using the LLM as an Illumination tool to organize, summarize and reify their existing thoughts, rather than for ideation. Once an idea could be articulated or formalized, participants often experimented with them by writing them down during the Implementation stage, either as a title while slogan writing or as a film script or storyline while story writing. In general, the three stages often occurred in the linear order, where creative thoughts were noticed during Ideation, elucidated during Illumination, and experimented with during Implementation. However, just as previous theories of writing suggest [25, 57], the process of prewriting is also iterative, notably the usage of LLMs for Ideation. We found that participants often used the LLM for Ideation whenever they encountered writer’s block during the Implementation stage. Furthermore, the unexpected results and randomness (sometimes even “failure”) of the LLM output from any of the three stages was also considered to be a source of inspiration, which implicitly led to Ideation.”