Scholarlink

<p>Innovation has long been regarded as a uniquely human capability; however, the rapid development of generative artificial intelligence, particularly large language models such as ChatGPT, is increasingly challenging this assumption. Taking environmental research as an example, this study evaluates the innovation-like behavior of ChatGPT through a proxy task of predicting future research hotspots. Fed with 20&#xa0;years of previous literature from a professional environmental journal, the optimal ChatGPT setup correctly predicted 80% of the hotspots in the next year, but the correct predictions highly relied on repeating existed keywords, reflecting the model’s ability to capture thematic continuity rather than genuine scientific innovation. Interestingly, sometimes it correctly predicted new words beyond the history hotspots list. The new words were found meaningful and hard to be deducted by existed keywords, which is seen as a weak signal of novelty generation. In conclusion, ChatGPT seems unable to substitute human beings in scientific innovation in its current state, but the capability it exhibited and the following ethics problems deserve to be carefully concerned in advance.</p> Graphical Abstract <p></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

Will artificial intelligence challenge human in innovative works?– a perspective in environmental research

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Yuxuan Chen,
  • Junyu Tao,
  • Guanyi Chen,
  • Beibei Yan,
  • Chao Chen,
  • Rui Liang,
  • Zhanjun Cheng,
  • Li’an Hou

摘要

Innovation has long been regarded as a uniquely human capability; however, the rapid development of generative artificial intelligence, particularly large language models such as ChatGPT, is increasingly challenging this assumption. Taking environmental research as an example, this study evaluates the innovation-like behavior of ChatGPT through a proxy task of predicting future research hotspots. Fed with 20 years of previous literature from a professional environmental journal, the optimal ChatGPT setup correctly predicted 80% of the hotspots in the next year, but the correct predictions highly relied on repeating existed keywords, reflecting the model’s ability to capture thematic continuity rather than genuine scientific innovation. Interestingly, sometimes it correctly predicted new words beyond the history hotspots list. The new words were found meaningful and hard to be deducted by existed keywords, which is seen as a weak signal of novelty generation. In conclusion, ChatGPT seems unable to substitute human beings in scientific innovation in its current state, but the capability it exhibited and the following ethics problems deserve to be carefully concerned in advance.

Graphical Abstract