🤖 AI

Large language models accurately predict public perceptions of support for climate action worldwide

This preregistered study demonstrates that large language models, particularly Claude, can accurately predict global public perceptions of support for climate action and the associated perception gaps using country-level indicators, offering a rapid and cost-effective alternative to traditional surveys while revealing a systematic downward bias in social projection.

Original authors: Nattavudh Powdthavee, Sandra J. Geiger

Published 2026-01-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nattavudh Powdthavee, Sandra J. Geiger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Silent Majority"

Imagine a massive party where almost everyone wants to clean up the mess. In reality, about 89% of people want the government to do more about climate change, and 69% are willing to chip in a little money to help.

But here's the catch: Nobody knows that.

Most people walk into that party thinking, "Oh, everyone else is probably just going to sit on the couch and do nothing." This is called pluralistic ignorance. Because everyone thinks everyone else doesn't care, no one speaks up or contributes. They are all waiting for someone else to start, creating a cycle of silence and inaction.

The New Tool: The "Super-Oracle"

For a long time, the only way to find out what people really think (and what they think others think) was to send out expensive surveys to thousands of people in every country. This is like hiring a team of detectives to interview every guest at the party. It takes forever and costs a fortune.

The authors of this paper asked a new question: Can Artificial Intelligence (specifically Large Language Models or LLMs) act as a "Super-Oracle" to guess these feelings without sending out a single survey?

They tested four famous AI models (GPT-4o mini, Claude 3.5 Haiku, Gemini 2.5 Flash, and Llama 4 Maverick) to see if they could predict how much people in 125 different countries underestimate each other's willingness to help.

The Results: The AI Got It Right (Mostly)

The researchers treated the AI like a student taking a test. They gave the AI real data from 125 countries (like GDP, internet usage, and actual survey results) and asked it to guess the "perception gap" (the difference between what people think others will do vs. what they actually will do).

The Star Performer: One model, Claude, was incredibly accurate. It guessed the size of the perception gap with an error margin of only about 5 percentage points. It was just as good as a traditional statistical computer model.
The Runner-Up: Llama also did a great job.
The Strugglers: The other two models (GPT and Gemini) were less accurate, often guessing that the gap was smaller than it really was.

The Analogy: Imagine trying to guess how much your neighbors are willing to pay for a new park. The best AI models looked at the neighborhood's income and demographics and said, "They think their neighbors won't pay much, but actually, they are willing to pay a lot." They got the math right.

How Did the AI Do It? (The "Magic" vs. The "Reasoning")

The researchers were worried the AI might just be cheating. Since these models are trained on huge amounts of text from the internet, they might have just "memorized" the answers from a specific survey that was published recently.

To test this, they played a game of "spot the difference":

The "Name Swap" Test: They told the AI, "This is France, but here are the economic stats of Nigeria." If the AI was just memorizing that "France = X," it would guess X. But the AI looked at the Nigerian stats and guessed based on those. This proved it was reasoning, not just reciting a script.
The "Missing Info" Test: They removed specific facts (like income levels) to see if the AI could still guess. The best models (Claude and Llama) realized that the most important clue was simply: "How willing are people to help themselves?" If people are willing to help themselves, they assume others are too. This is a psychological concept called social projection.

The Conclusion: The AI wasn't just a dictionary looking up answers; it was acting like a smart psychologist, using logic to figure out how humans think about other humans.

Where Does the AI Work Best?

The AI isn't equally good everywhere.

The "High-Tech" Zone: In wealthy countries with high internet use (like Europe and North America), the AI was very accurate. It's like a detective who has access to a massive library of books about those specific places.
The "Low-Tech" Zone: In poorer countries or places with less internet access, the AI was less accurate. This is because the AI was trained mostly on data from rich, connected countries. It's like a detective who has never visited a remote village and has to guess based on stereotypes.

What Does This Mean?

The paper concludes that AI can be a fast, cheap "thermometer" for measuring these perception gaps around the world.

In rich countries, it can replace expensive surveys.
In poorer countries, it can give a "rough estimate" to tell us where the problem is worst, so we know where to send real human surveyors later.

Crucially, the paper does not claim that AI can fix climate change or that we should stop doing real surveys entirely. It simply says AI is a powerful new tool to help us see the "invisible" gap between what we think and what is actually true, helping us know where to start the conversation.

Technical Summary: Large Language Models Accurately Predict Public Perceptions of Support for Climate Action Worldwide

Problem Statement

Addressing climate change requires coordinated shifts in individual behavior and systemic policy. While global opinion polls indicate broad support for climate action (e.g., 69% willingness to contribute 1% of income), actual mitigation efforts remain insufficient. A primary barrier identified in multidisciplinary research is pluralistic ignorance: individuals systematically underestimate the level of support and willingness to act among their peers. These misperceptions (second-order beliefs) suppress expressed support and action through conditional cooperation mechanisms, creating a self-perpetuating spiral of inaction.

Correcting these gaps requires timely, global maps of perception discrepancies. However, generating such data is logistically challenging and costly, as it necessitates paired measures of personal and perceived attitudes from representative samples across diverse nations. This study investigates whether Large Language Models (LLMs) can serve as a scalable, cost-effective alternative or complement to traditional surveys for mapping these perception gaps worldwide.

Methodology

Data and Ground Truth

The study utilizes data from the Global Climate Change Survey (part of the Gallup World Poll 2021/22), covering 125 countries with approximately 129,902 respondents.

Outcome Variable: The "perception gap," defined as the difference between the country-level average of personal willingness to contribute financially to climate action (first-order belief) and the perceived willingness of others (second-order belief).
Covariates: Country-level sociodemographic (age, education, religiosity), macroeconomic (GDP per capita, inequality, HDI), and climate-related (average temperature) indicators were aggregated from the World Bank and Gallup data.

Experimental Design

The authors conducted a preregistered experiment (OSF: ez4ur) comparing four state-of-the-art LLMs against statistical baselines:

Models Evaluated: GPT-4o mini, Claude 3.5 Haiku, Gemini 2.5 Flash, and Llama 4 Maverick (17B).
Prompting Protocol: Models were queried under eight sequential information stages (Stage 1 to Stage 8), ranging from providing only the country name to providing the full suite of sociodemographic, economic, climate, and first-order willingness data.
Task: Models were instructed to estimate the percentage of respondents in a given country who believe others are willing to contribute 1% of their income to climate action, based strictly on the provided contextual data.
Benchmarks: LLM performance was compared against:
- Ground Truth: The actual survey-derived perception gaps.
- Statistical Models: Ordinary Least Squares (OLS) and Lasso regressions trained on 80% of countries and evaluated on a 20% holdout set.

Diagnostic and Mechanistic Tests

To distinguish between structured inference and memorization of training data, the authors employed several diagnostic tests:

Feature Ablation: Systematically removing specific input variables (e.g., removing personal willingness, GDP, or demographics) to measure the impact on prediction error.
Counterfactual Perturbations:
- Willingness Flip: Swapping high/low personal willingness values between countries.
- Name Mismatch: Assigning a country's name to a different country's data profile.
- GDP Flip: Inverting GDP values across countries.
Robustness Checks: Testing across different prompt formats (bullet points vs. sentences) and generalizing to 24 alternative U.S. climate policy perception gaps.

Key Results

Predictive Accuracy

Top Performers: Claude 3.5 Haiku and Llama 4 Maverick demonstrated the highest accuracy.
- Claude: Achieved a Mean Absolute Error (MAE) of 4.60 percentage points (p.p.) and a correlation ( $r$ ) of 0.77 with ground truth, performing comparably to OLS (MAE 4.79) and Lasso (MAE 4.66) models.
- Llama: Showed similar performance with an MAE of 7.00 p.p. and $r = 0.67$ .
Lower Performers: GPT-4o mini and Gemini 2.5 Flash exhibited significantly higher errors (MAE > 13 p.p. and 14 p.p., respectively) and lower correlations. Gemini's performance was particularly poor ( $r = 0.15$ ), likely due to safety filtering or model-specific limitations.
Global Pattern: All models captured the broad spatial structure of perception gaps, with underestimations being most severe in Latin America, West Africa, and South Asia, and smaller in Western Europe.

Structural Predictors of Accuracy

Prediction accuracy was not uniform; it correlated strongly with structural indicators of information availability:

Internet Penetration & GDP: Models performed significantly better in countries with higher internet penetration and GDP per capita ( $R^2$ ranges 0.08–0.27).
Bias: This suggests a representation bias in training data, where LLMs are more reliable in resource-rich, digitally connected nations and less accurate in poorer or less connected settings.

Mechanistic Insights

Social Projection: The best-performing models (Claude, Llama) correctly identified personal willingness (first-order belief) as the strongest predictor of perceived willingness (second-order belief), mirroring the human psychological process of social projection.
Reasoning vs. Memorization:
- Ablation: Removing personal willingness data increased errors for Claude and Llama by ~1.5–2.2 p.p., confirming reliance on this feature.
- Counterfactuals: Swapping country names or GDP values had negligible effects on the top models, indicating they rely on structural reasoning of the provided features rather than retrieving memorized country-specific statistics.
- Training Data Exposure: Although some models were trained after the survey data was released, the pattern of errors (varying by information environment, not uniform) and the failure of the most recently trained model (Gemini) to perform best argue against pure memorization.

Significance and Claims

The paper claims that LLMs, particularly Claude and Llama, provide a rapid, cost-effective tool for assessing global perception gaps regarding climate action. The significance of these findings is framed as follows:

Scalability: LLMs can map perception gaps across 125 countries with accuracy comparable to traditional statistical models, offering a viable alternative to expensive, logistically difficult large-scale surveys.
Targeted Intervention: By identifying where perception gaps are largest (e.g., specific regions in Africa and Asia), LLMs can help governments and NGOs prioritize where misperception-correction efforts are most urgently needed and likely to be effective.
Complementary Role: The authors modestly position LLMs not as a total replacement for surveys, but as a complementary mechanism. They can enable cost-effective tracking in information-rich environments and provide provisional signals in resource-scarce settings, guiding where direct measurement resources should be deployed.
Mechanistic Validity: The study demonstrates that LLMs can engage in complex second-order reasoning ("I think that you think") and capture the psychological mechanism of social projection with systematic downward bias, provided they are conditioned on relevant structural and attitudinal data.

The authors conclude that while limitations exist regarding subnational granularity and performance in low-connectivity regions, LLM-based perception mapping represents a significant advancement in computational social science for climate policy design.