Evaluation of Large Language Models as Tools, Models, and Partners in Creative Thinking Research: A Selective Narrative Review with the GCA Framework

Abstract

Creativity research faces three persistent bottlenecks: divergent-thinking scoring is labour-intensive, cognitive models of creativity remain underspecified, and laboratory tasks fall short of real-world creative achievement. Large language models (LLMs) offer potential solutions, but the field lacks a structured framework for evaluating them. This selective narrative review (January 2018–June 2026) applies the generation–capability–assessment (GCA) framework, whose three axes are operationalised through descriptive criteria with provisional heuristic thresholds. On the generation axis, LLMs exceed average human performance on divergent-thinking tasks in most independent comparisons (Hedges’ g ≈ 0.5–2.6), an advantage qualified by fluency dependency, the superiority of top-performing humans at scale, and a novelty–typicality trade-off. On the capability axis, LLMs simulate some task-level associative behaviour and can generate hypotheses for human research, but there is no evidence that they instantiate human-like creative mechanisms. On the assessment axis, automated scoring shows promising reliability and convergent validity for specific languages and tasks (ICC ≥ 0.80 and r ≥ 0.70 in selected studies), but cross-language generalisation is largely untested and individual-level use is unsupported. Human–AI co-creativity may benefit from a division of labour, although social–affective dimensions may matter more than cognitive support. We provide a GCA reporting protocol and identify research priorities.