EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Can large language models predict public perceptions of support for climate action across 125 countries?

Nattavudh Powdthavee iD, Sandra J Geiger iD

PNAS Nexus · 30.10.2026 · 5 (11)

DOI: 10.1093/pnasnexus/pgag374 ↗

Kurz gefasst

Fachartikel in der Zeitschrift PNAS Nexus, online erschienen am 31.10.2026. Verfasst von Nattavudh Powdthavee und 1 weiteren Autor:innen. Beteiligte Einrichtung: Division of Economics, Nanyang Technological University, 48 Nanyang Avenue, Singapore 639798, Singapore. Die Arbeit nennt 67 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Second-order beliefs (“what people think others think”) are theorized to be a powerful driver of climate action, yet they remain costly and difficult to measure at a global scale. Here, we tested whether large language models (LLMs) can predict second-order beliefs about willingness to contribute financially to address climate change across 125 countries. In a preregistered experiment, we prompted four contemporary LLMs and benchmarked their predictions against nationally representative samples from the Gallup World Poll 2021/2022 and statistical regression models. Among the LLMs, Claude and Llama achieved the lowest average errors (∼5 and 7 percentage points, respectively). Claude's error was not statistically different from those of ordinary least squares and Lasso, whereas Llama, GPT, and Gemini performed significantly worse. Yet, accuracy depended strongly on whether prompts included each country's actual first-order willingness to act. Without it, correlations between predicted and observed second-order beliefs were close to zero across all models, even when demographic, economic, and climate information was provided. In exploratory analyses, Claude estimated first-order willingness from country-level information and used these estimates to predict second-order willingness; however, the resulting second-order predictions were unrelated to the observed values. Counterfactual and memorization checks found little evidence that the models were retrieving memorized values; the models appear mainly to map the supplied first-order value onto a second-order estimate, with little contribution from the other country-level information. Overall, the tested approach cannot substitute for survey measurement in settings where first-order data are unavailable.

Schlagwörter

Stichwörter der Autor:innen

climate change perceptionscomputational social sciencelarge language modelspluralistic ignorancesecond-order beliefssocial projection

Autor:innen und Institutionen

  • Division of Economics, Nanyang Technological University, 48 Nanyang Avenue, Singapore 639798, Singapore.
  • Behavioral Science for Policy Lab, Princeton University, Princeton, NJ 08540, USA.

ORCID-Profile: Nattavudh Powdthavee, Sandra J Geiger

Zitieren

APA

Powdthavee, N., & Geiger, S. J. (2026). Can large language models predict public perceptions of support for climate action across 125 countries?. PNAS Nexus, 5(11). https://doi.org/10.1093/pnasnexus/pgag374

Vancouver

Powdthavee N, Geiger SJ. Can large language models predict public perceptions of support for climate action across 125 countries?. PNAS Nexus. 2026;5(11). doi:10.1093/pnasnexus/pgag374

BibTeX
@article{Powdthavee2026,
  author = {Powdthavee, Nattavudh and Geiger, Sandra J},
  title = {Can large language models predict public perceptions of support for climate action across 125 countries?},
  journal = {PNAS Nexus},
  year = {2026},
  volume = {5},
  number = {11},
  doi = {10.1093/pnasnexus/pgag374}
}
Kostenlos starten →