Existing qualitative studies suggest that GenAI’s strongest value is upstream: ideation and exploratory prototyping, not end-to-end authorship. Across studies, the shared conclusion is that systems broaden option spaces but humans still frame, judge, and steer what becomes “the work”

A consistent finding across the corpus is that human-in-the-loop refinement is the production norm. Generative outputs behave like provisional artefacts: teams iterate, curate, correct, and integrate, and prompting becomes progressive specification work rather than a one-shot query.

(4/) The story gets interesting with individual differences: Amongst others, AI literacy moderated the effect with lower literacy -> process tended to lower creativity ratings; higher literacy -> process tended to raise them.

(3/) Prolific participants were UK census-representative (N=298+295). The headline Study 1 result: process visibility did NOT increase perceived creativity on average. Study 2 sought to “help” the process animation by adding a tutorial on diffusion: still no tutorial/PE/interaction effects!

3/5 More specifically, we demonstrate that modifying the SO model learning parameters gives rise to four different regimes that can account for both creative products and inconclusive outcomes, thus providing a framework for studying and understanding the creative potential of learning systems.

2/5 Developed to model complex adaptive systems in ALife and advocated as a candidate for minimal agency, the Self-Optimization (SO) model can be considered as the 3rd operational mode of the classical Hopfield Network, leveraging the power of associative memory to enhance optimization performance.