Scholarlink

<p>Early large language models (LLMs) were released with minimal alignment, offering a rare view of how generative systems reframe the ethical values embedded in human texts. We examine outputs from a 2021 version of OpenAI’s base GPT-3, prompting it to summarise culturally diverse source materials (laws, political speeches, and philosophical works) and interpreting results through a descriptive, moral value pluralist lens. Where possible, we contextualise outputs with cross-national datasets such as the World Values Survey. We document recurring value drift: Australia’s firearm policy is recast as a threat to liberty; de Beauvoir’s feminist critique becomes gender-essentialist dating advice; and Merkel’s humanitarian appeal is recast as immigration control. In contrast, multilateral documents (UN/UNESCO) exhibit greater value stability, suggesting consensus-crafted language can buffer against cultural mutation. We argue that these early behaviours (observed before extensive fine-tuning and safety layers) provide a baseline for understanding how training distributions shape normative framing. Our contribution is twofold: (1) empirical evidence that value drift can invert or overwrite encoded values along predictable cultural axes, and (2) a pluralist, descriptive evaluation method that surfaces whose values dominate and when. We conclude with implications for culturally inclusive evaluation and alignment in contemporary LLMs.</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

The ghost in the machine speaks with an American accent: cultural value drift in early GPT-3 and the case for pluralist evaluation of generative AI

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月23日

查看全文
  • Rebecca Johnson,
  • Leslye Denisse Dias Duran,
  • Enrico Panai,
  • Natalia Menéndez González,
  • Giada Pistilli,
  • Julija Kalpokienė,
  • Donald Jay Bertulfo

摘要

Early large language models (LLMs) were released with minimal alignment, offering a rare view of how generative systems reframe the ethical values embedded in human texts. We examine outputs from a 2021 version of OpenAI’s base GPT-3, prompting it to summarise culturally diverse source materials (laws, political speeches, and philosophical works) and interpreting results through a descriptive, moral value pluralist lens. Where possible, we contextualise outputs with cross-national datasets such as the World Values Survey. We document recurring value drift: Australia’s firearm policy is recast as a threat to liberty; de Beauvoir’s feminist critique becomes gender-essentialist dating advice; and Merkel’s humanitarian appeal is recast as immigration control. In contrast, multilateral documents (UN/UNESCO) exhibit greater value stability, suggesting consensus-crafted language can buffer against cultural mutation. We argue that these early behaviours (observed before extensive fine-tuning and safety layers) provide a baseline for understanding how training distributions shape normative framing. Our contribution is twofold: (1) empirical evidence that value drift can invert or overwrite encoded values along predictable cultural axes, and (2) a pluralist, descriptive evaluation method that surfaces whose values dominate and when. We conclude with implications for culturally inclusive evaluation and alignment in contemporary LLMs.