Do Large Language Models Show a Human-Like Serial Order Effect in Divergent Creativity Tasks? Evidence from the Alternative Uses Task

This paper concerns the behaviour of widely used Large Language Models (LLMs) in a prototypical creative task, viz., the Alternative Uses Task (AUT). Previous studies have indicated that LLMs’ average performances on the AUT are at or somewhat above average human levels but have not examined patterns within performance data (such as the Serial Order Effect (SOE)) which could suggest similarities and differences between human and LLM divergent production processes. The present exploratory study tested for and found evidence for human-like SOEs in AUT performance across a range of LLMs.