Scholarlink

<p>Recent advances in generative artificial intelligence (GenAI) have led to their use in a wide range of professions and tasks requiring creativity. Yet, there is no literature on how to optimally elicit creative output from GenAI chatbots. We systematically evaluated nine GenAI chatbots using a divergent thinking task (i.e., Alternate Uses Task) to explore whether the originality of GenAI ideas differs within a single response and across repeated prompts for more ideas. We analyzed the 1250 responses for each chatbot using repeated-measures ANOVA, which showed no significant effect of response order on idea originality for all GenAI models (<i>p</i> &gt; 0.05). Similarly, repeated prompting for more ideas generally did not affect the originality of the ‘average’ or ‘best’ ideas. The sole exception was DeepSeek-V3, which exhibited meaningfully higher originality scores after being prompted twice for more ideas (<i>p</i> = 0.01). These findings suggest that, in contrast to humans, GenAI models maintain highly stable ideational patterns, which is useful for informing best practices for interpreting GenAI outputs in creative workflows. We also discuss the implications of human-AI co-creativity and avenues for future research.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

How to maximize the creativity of artificial intelligence: an experimental analysis of response order and prompting effects

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Vikram Arora,
  • Adam Sutoski,
  • Joelle Boilard,
  • Elyssia Leone,
  • Hannah Zizzo,
  • Borong Wang,
  • Goran Calic,
  • Sameer Parpia,
  • Mohit Bhandari,
  • Alex Thabane

摘要

Recent advances in generative artificial intelligence (GenAI) have led to their use in a wide range of professions and tasks requiring creativity. Yet, there is no literature on how to optimally elicit creative output from GenAI chatbots. We systematically evaluated nine GenAI chatbots using a divergent thinking task (i.e., Alternate Uses Task) to explore whether the originality of GenAI ideas differs within a single response and across repeated prompts for more ideas. We analyzed the 1250 responses for each chatbot using repeated-measures ANOVA, which showed no significant effect of response order on idea originality for all GenAI models (p > 0.05). Similarly, repeated prompting for more ideas generally did not affect the originality of the ‘average’ or ‘best’ ideas. The sole exception was DeepSeek-V3, which exhibited meaningfully higher originality scores after being prompted twice for more ideas (p = 0.01). These findings suggest that, in contrast to humans, GenAI models maintain highly stable ideational patterns, which is useful for informing best practices for interpreting GenAI outputs in creative workflows. We also discuss the implications of human-AI co-creativity and avenues for future research.