arXiv cs.CY· Sola Kim, Jieshu Wang, Marco A. Janssen, John M. Anderies·· 9 小时前AI 评分34
大语言模型如何歪曲美国气候舆论
How Large Language Models Misrepresent American Climate Opinions
AI 导读
研究用978名美国气候舆论调查受访者的画像提示六个大语言模型,对比20道题的生成回答后发现,模型将预测拉向人群均值,预测回答方差平均比人类低29%。模型对性别的假设在各族裔中一致,却与黑人美国人的实际模式方向相反;仅审计性别会误判偏差方向,需交叉性审计才能发现。
正文 · 原文
Abstract:Large language models are increasingly used to simulate public opinion, standing in for survey respondents as "silicon samples." Whether they faithfully represent U.S. climate opinion, particularly at the intersections of race, gender, and ideology, remains largely untested. We prompted six large language models with the profiles of 978 respondents to a nationally representative U.S. climate opinion survey and compared the generated responses with those respondents' own answers across 20 questions. The models compress the diversity of American climate opinion, pulling predictions toward the population mean: less-concerned groups are predicted as more concerned, more-concerned groups as less, and the variance of predicted responses falls, on average, 29% below that of human responses. The compression is intersectional. The models apply one gender assumption, predicting women as more pro-climate than men in every racial group. The assumption matches the direction of the human pattern for White, Hispanic, and Other Americans and runs the wrong way for Black Americans. We claim the direction of that difference rather than its size, since it rests on small cells at the ideological extreme. An audit of gender alone would report that the models understate the gender gap, and would miss that the error changes direction by race, which is what single-variable audits cannot detect. What is at stake is not accuracy alone but representation: whose views a synthetic sample renders legible.
| Comments: | 45 pages, 11 figures. Accepted for publication in Technology in Society. v3: a normalization correction scales items by instrument range rather than observed range, revising the reported variance compression from 27.9% to 29% and all coefficient units. Five appendices added in review: decoding sensitivity, matched response space, don't-know handling, item missingness, small-cell robustness |
| Subjects: | Computers and Society (cs.CY); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2512.23889 [cs.CY] |
| (or arXiv:2512.23889v3 [cs.CY] for this version) | |
| https://doi.org/10.48550/arXiv.2512.23889 arXiv-issued DOI via DataCite |
Submission history
From: Sola Kim [view email]
[v1]
Mon, 29 Dec 2025 22:29:10 UTC (108 KB)
[v2]
Sat, 18 Apr 2026 07:30:36 UTC (74 KB)
[v3]
Thu, 19 Nov 2026 16:42:14 UTC (106 KB)
来源:arXiv cs.CY · arxiv.org