Scholarlink

<p>The emerging linguistic capabilities of neural language models (NLM) sparked interest in applying evaluation methodologies traditionally reserved for humans. From a machine psychology perspective, this study examines OpenAI GPT-3.5’s ability to critically reason about moral issues, comparing it to a human sample. A sample of 480 data points generated by neural language models (NLMs) and 480 human subjects, selected based on academic scores and gender, completed the Critical Reasoning Assessment (CRA). The results revealed a higher capacity for critical thinking in NLMs compared to humans. In alignment with the human sample, the artificial sample showed a positive correlation between academic score and critical reasoning ability, with no significant differences related to gender. These findings broaden our understanding of reasoning in NLMs and suggest that models like GPT-3.5 might surpass humans in critical reasoning. These discoveries are crucial for the development of ethical artificial intelligences and the practical application of their capabilities in various contexts that require problem-solving and decision-making on complex issues.</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

Do neural language models critically reason about moral issues like humans?

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Rossella Suriano,
  • Alessandro Acciai,
  • Rosa Angela Fabio,
  • Lucia Guerrisi,
  • Alessio Plebe

摘要

The emerging linguistic capabilities of neural language models (NLM) sparked interest in applying evaluation methodologies traditionally reserved for humans. From a machine psychology perspective, this study examines OpenAI GPT-3.5’s ability to critically reason about moral issues, comparing it to a human sample. A sample of 480 data points generated by neural language models (NLMs) and 480 human subjects, selected based on academic scores and gender, completed the Critical Reasoning Assessment (CRA). The results revealed a higher capacity for critical thinking in NLMs compared to humans. In alignment with the human sample, the artificial sample showed a positive correlation between academic score and critical reasoning ability, with no significant differences related to gender. These findings broaden our understanding of reasoning in NLMs and suggest that models like GPT-3.5 might surpass humans in critical reasoning. These discoveries are crucial for the development of ethical artificial intelligences and the practical application of their capabilities in various contexts that require problem-solving and decision-making on complex issues.