New Study: Many LLMs Express Greener Attitudes Than Average Humans

A new study from researchers at the University of Bremen and colleagues evaluates the environmental attitudes embedded in 31 widely used proprietary and open-weight large language models. By adapting questions from established environmental awareness surveys and adding behavioral measures, the team compared LLM responses against human survey benchmarks from Germany. The results show that many models—including popular systems from OpenAI, Meta, and others—express environmental cognition and affect that are more progressive than the average human respondent, and they recommend behaviors with higher potential CO2 reduction. However, the study found no systematic link between model origin, size, or release date and the degree of environmental friendliness.

The critical finding is that LLMs exhibit strong contextual sensitivity: when given a persona prompt (e.g., “you are a conservative voter”), their environmental attitudes shift dramatically toward the specified ideology. This sycophantic behavior raises serious concerns about reliability when LLMs are deployed for sustainability reporting, decision support, or public communication. The authors provide a reusable evaluation framework and benchmark dataset to help auditors assess value alignment in LLMs. They stress the need for governance and transparency as AI systems become embedded in sustainability transformations.