Large language models are being woven into decision-making across society, including choices with ecological consequences. A new preprint on arXiv, 'Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?', asks whether those environmentally friendly decisions remain robust when the model is under pressure. The abstract notes that models have biases, and that these are difficult to measure, particularly in closed-weight systems where the internal parameters are not publicly available.
Because the full paper is not yet available beyond the abstract, the specific methods and results are not described in the source. The key contribution, according to the abstract, is to highlight the measurement problem: without access to model internals, external audits must rely on behavioral tests. The title's reference to 'greenwash' suggests the paper examines whether LLMs merely appear to make ecological choices while failing under stress.
The abstract does not describe solutions, but it underscores a growing concern for regulators and users who need to trust LLM outputs in high-stakes domains. As closed-weight models become more common, the difficulty of verifying their decision-making biases will likely become a central issue.