Answers are grounded in this paper and cite their sources.
Large language models (LLMs) are increasingly proposed for sustainability assessment and decision support, yet their judgments may depend on the selected model and on how the task is defined. We compare 13 locally executed open-weight LLMs on 80 author-written sustainability scenarios using three prompt families that ask which actor should be responsible, bear the costs, or initiate action. Each model answered at two temperature settings with 100 repetitions, producing 624,000 deposited outputs, of which 610,708 contain a complete score vector. Scores were analyzed as relative actor support after removing each response’s overall level. Within this instrument, the cost prompt family produced the highest cross-model disagreement, exceeding both responsibility and catalyst by 7.37 points (question-cluster 95% intervals 5.56−9.15 and 5.38−9.36, describing variation across these 80 scenarios). It also produced the most lead-actor disagreement and the lowest full consensus. Moving from catalyst to cost reduced relative support for government and increased it for individuals and firms; moving from cost to responsibility largely reversed this pattern. These differences exceeded temperature-setting differences and persisted under model exclusion, alternative score transformations and tie rules, missing-response checks, and stems without explicit cost wording. Exploratory category patterns are conditional on uneven item coverage. Secondary rationale-language comparisons show the same within-model ordering but not the most cross-model variation under cost. Operational use should disclose model, prompt, and temperature choices; compare plausible alternatives; repeat generations; and refer high-disagreement judgments for human deliberation.
No takes yet. Share an insight, caveat, or question.
Olah et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: