LLMs Reinforce Status Quo Climate Advice, Revealing Structural Bias in AI Training Data Across Cultures
Original framing: “AI chatbots endorse existing climate-related policies twice as often as new proposals” — Phys.org
Indigenous climate knowledge, such as Māori concepts of kaitiakitanga (guardianship) and Andean ayllu reciprocity, offers policy alternatives centered on relational stewardship rather than market‑based mechanisms. These perspectives are systematically excluded from the corpora used to train LLMs, resulting in a bias toward technocratic, incremental solutions. Restoring this knowledge would expand the policy imagination of AI systems beyond the status quo.
The bias of LLMs toward existing climate policies is not a random error but a symptom of deeper epistemic injustices: training data privileging dominant discourses, historical patterns of technology reinforcing power, and the systematic exclusion of Indigenous, Global South, and marginalized knowledge traditions.
By recognizing AI as a trickster that both reveals and conceals alternatives, we can deploy deliberate interventions—diversifying data, auditing outputs, incentivizing novelty, and governing through inclusive boards—to shift AI from a status‑quo amplifier to a catalyst for transformative, justice‑oriented climate policy. Such systemic change aligns technological development with pluriversal visions of sustainability, ensuring that the tools we build serve the many rather than the few.