Silenced by design: Marginal voices and algorithmic bias in AI-generated climate discourse

Abstract

This research examines how large language models (LLMs) shape climate crisis discourse, often reproducing dominant, Global North-centric perspectives while marginalising alternative or justice-oriented voices, thereby facilitating climate misinformation—by reinforcing narrow epistemic frameworks that exclude correction from diverse knowledge communities—if left unchecked. Drawing on ecolinguistics and posthumanist thought, it analyses AI-generated responses to climate justice prompts, using Critical Discourse Analysis (CDA) and framing analysis. In so doing, the study reported herein sets out to explore tone, framing, agency construction, epistemic stance and omission/bias with a view to identifying the discursive mechanisms through which LLMs reproduce structural silences and marginalise justice-oriented climate epistemologies. Findings from six chatbots (GPT-3.5, GPT-4o, Claude, Gemini, Copilot, Perplexity) reveal that inclusive framings emerge mainly when directly prompted; when left to their defaults, instead, their outputs visibly privilege technocratic and institutional narratives over plural, situated and relational ecologies. The study claims that such defaults carry real epistemic costs, as they perpetuate exclusion and limit public engagement with diverse environmental knowledges, which highlights the need for more justice-sensitive design and critical AI literacy development.