A Multilingual Investigation of Anthropocentrism in GPT-4o

This paper presents a methodology to assess anthropocentric bias in Large Language Model (LLM)-generated content (GPT-4o) across languages. Anthropocentric bias refers to the systematic prioritization of human perspectives, needs, and values over those of non-human entities, often resulting in language that marginalizes or instrumentalizes the natural world. Using a multilingual setup involving English, Italian, and German, we prompted the model with 150 inputs across three ideologically framed conditions (neutral, anthropocentric, ecocentric). Following an exploratory phase and prompt refinement, we analysed the model’s responses through noun phrases and verbs. As a second contribution, we release a manually curated multilingual glossary of 424 ecologically relevant noun phrases, provided as an open resource to support future ecocritical analyses. In our quantitative and qualitative analysis, we examined how non-human entities are framed, what verbs and connotations are associated with them, and how these patterns vary across prompts and languages. Results show that anthropocentric framing systematically emerges even in neutral and ecocentric outputs, with notable cross-linguistic differences, suggesting that such bias is structurally embedded in the model’s behaviour.