Scholarlink

<p>Human–machine partnerships are increasingly used to address grand societal challenges, yet knowledge of the comparative strengths of humans and machines to innovate is nascent. Here we compare the ability of humans (<i>N</i> = 9,198) and large language models (LLMs, <i>N</i> = 215,542 observations) to generate novel ideas in an established creativity task. We present three key results. First, human creativity on average is slightly higher than that of LLMs. Second, creativity differences are pronounced at the extremes of the distribution, with humans exhibiting greater variability and higher levels of creativity in the right-hand tail of the distribution. Third, attempts to increase the creativity of LLMs through instructing LLMs to take on genius personas or different demographic roles lifted performance up to a threshold beyond which the output became opposite real-life patterns, whereas strategic prompt-engineering efforts yielded mixed to negative results. We discuss the implications of our findings for human–machine collaboration and problem solving.</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

A large-scale comparison of divergent creativity in humans and large language models

  • 期刊论文

  • 发表日期：

    2025年12月23日

查看全文
  • Dawei Wang,
  • Difang Huang,
  • Haipeng Shen,
  • Brian Uzzi

摘要

Human–machine partnerships are increasingly used to address grand societal challenges, yet knowledge of the comparative strengths of humans and machines to innovate is nascent. Here we compare the ability of humans (N = 9,198) and large language models (LLMs, N = 215,542 observations) to generate novel ideas in an established creativity task. We present three key results. First, human creativity on average is slightly higher than that of LLMs. Second, creativity differences are pronounced at the extremes of the distribution, with humans exhibiting greater variability and higher levels of creativity in the right-hand tail of the distribution. Third, attempts to increase the creativity of LLMs through instructing LLMs to take on genius personas or different demographic roles lifted performance up to a threshold beyond which the output became opposite real-life patterns, whereas strategic prompt-engineering efforts yielded mixed to negative results. We discuss the implications of our findings for human–machine collaboration and problem solving.