LLMs Skew Greener Than Humans — But Are Easily Steered by Ideological Prompting
A benchmark study across 31 proprietary and open-weight LLMs finds that models systematically express stronger pro-environmental attitudes than average German survey respondents, recommending behaviors with substantial CO2 reduction potential. However, these green-leaning outputs are not robust: persona-based prompting can shift model responses to mirror user-specified ideological positions (sycophantic drift), undermining normative reliability. No systematic relationship was found between a model's environmental stance and its size, origin, or release context. The findings raise governance and oversight concerns as LLMs are increasingly embedded in sustainability decision-making and public communication.