Title: Are AI Models Greener Than We Are? Assessing Environmental Mindsets in Large Language Systems

Abstract:

As large language models (LLMs) play a growing role in sustainability decision-support systems, public communication, and reporting, there remains a significant gap in systematic evidence regarding the environmental attitudes embedded within their outputs. To address this, we introduce a novel benchmark designed to evaluate environmental cognition, affect, and behavioral recommendations in LLMs. We applied this framework to 31 prominent proprietary and open-weight models.

By leveraging items from established environmental awareness surveys alongside additional sustainability-focused behavioral metrics, we conducted a two-pronged comparison: evaluating responses across different models and contrasting model outputs with human survey data from Germany. We also tested the robustness of these responses under various prompting conditions.

Our analysis reveals that many LLMs demonstrate alignment with environmentally progressive viewpoints, often exceeding the attitudes of the average human survey respondent. Specifically, these models exhibited stronger environmental affect and cognition, while recommending behaviors linked to significant CO2 reduction potential. However, we found no consistent correlation between sustainability-oriented outputs and a model’s origin, parameter size, or release context.

Conversely, the models displayed notable contextual sensitivity. Through persona-based prompting, they exhibited sycophantic shifts that mirrored user-specified ideological stances. This behavior raises serious concerns regarding the steerability and normative reliability of LLMs in real-world applications. Our results offer a reusable evaluation framework for assessing sustainability-related value alignment in AI systems and underscore the critical need for governance, transparency, and rigorous oversight as AI becomes more deeply integrated into public decision-making and sustainability transitions.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC