Cultural value drift in early GPT-3 and the case for pluralist evaluation of generative AI
Published in Springer Nature’s AI and Ethics, this paper began with a deceptively simple question: when generative AI summarises a text, is it preserving the original values or quietly rewriting them?
Using one of the earliest studies of GPT-3, we found that the model sometimes recast culturally specific texts through dominant American moral frames. Australian firearms legislation became a story about individual liberty. Simone de Beauvoir’s feminist critique became dating advice. Humanitarian arguments about refugee protection shifted towards immigration control. In these cases, the facts were often preserved, but the values drifted.
The findings challenged a common assumption that generative AI is culturally neutral. They suggested that models can amplify some value systems while muting others, raising questions about representation, fairness, and whose perspectives become embedded in AI systems at global scale.
This visual explainer summarises the paper’s central findings and reflects on why they still matter today. As generative AI becomes increasingly integrated into institutions, products, and public life, evaluation must move beyond accuracy and safety alone. We also need ways to examine how values travel, shift, and are reproduced through AI systems operating across diverse cultural contexts.
Published on LinkedIn in April 2026