Fairness in LLM-Generated Surveys

Technical Summary: Fairness in LLM-Generated Surveys

Problem Statement
Large Language Models (LLMs) have emerged as viable alternatives to traditional surveys for simulating socio-political and economic patterns, capable of generating synthetic datasets that mirror human responses. However, the global applicability of these models is questionable due to unexplored biases stemming from their training data, which is predominantly U.S.-centric. While LLMs demonstrate high performance in well-represented contexts like the United States, their ability to accurately and fairly model socio-political behaviors in underrepresented regions, such as Chile, remains underexplored. This study addresses the critical gap in understanding how LLMs perform across diverse socio-demographic and geographic contexts, specifically examining whether these models perpetuate existing societal biases or inadequately represent specific populations.

Methodology
The study employs a comparative framework utilizing public survey data from Chile (Centro de Estudios Públicos - CEP) and the United States (American National Election Studies - ANES). The research design includes three prediction experiments:

1. Presidential Voting: Predicting vote choices in the 2020 U.S. election and the 2021 Chilean presidential election.
2. Abortion Opinion: Predicting individual stances on abortion based on socio-demographic characteristics.
3. Constitutional Plebiscite: Predicting outcomes for the 2022 Chilean constitutional referendum.

Models and Prompts: The study evaluates multiple models, including GPT-3.5-turbo, GPT-4-turbo, Llama-13b, and Mistral (v0.2). A "default prompt" configuration is used, incorporating Chain-of-Thought (CoT) reasoning, few-shot examples (5 random instances), and English as the primary language, though Spanish is also tested. To ensure fair cross-country comparison, an in-sample Random Forest is trained to establish an upper bound on prediction accuracy, serving as a baseline to normalize model performance against inherent data predictability.

Metrics:

• Accuracy: Measures individual-level prediction correctness.
• Jensen-Shannon Similarity (JSS): Measures the global alignment between predicted and actual probability distributions (group-level patterns).
• Overall Accuracy Equality: A fairness criterion evaluating if prediction accuracy is equal across protected and unprotected groups.
• Logistic Regression: Used to analyze how specific socio-demographic variables (gender, age, education, political identity, etc.) and their intersections influence the likelihood of a correct prediction.

Key Contributions

1. Novel Framework: Introduction of a framework for measuring socio-demographic biases in LLMs across different geographic contexts.
2. Comparative Evaluation: A direct comparison of LLM predictive capabilities using comparable political and moral questions from Chile and the U.S.
3. Bias Mitigation Analysis: An examination of prompting techniques and fine-tuning strategies to address performance disparities.

Results

• Geographic Disparities: Models consistently outperform on U.S. datasets compared to Chilean datasets across all metrics (Accuracy and JSS). This gap persists even after normalizing for dataset predictability, indicating a strong bias toward U.S.-centric training data.
• Model Performance: Llama-13b emerged as the best-performing model, often surpassing the in-sample Random Forest benchmark in capturing population-level trends (high JSS), though it still struggled with individual-level accuracy in Chile. Fine-tuning on Chilean data yielded only marginal improvements and, in some cases, introduced distributional biases.
• Socio-Demographic Biases:
  • Chile: Significant biases were observed against women, older adults, individuals with lower education levels, and religious individuals. Political identity also played a major role, with centrist views being less accurately predicted than strong left or right ideologies. Intersectional analysis revealed that senior, religious women with low education levels faced the lowest prediction accuracy.
  • United States: Performance was more stable across gender and age but showed disparities based on political identity (left-leaning views predicted more accurately) and income (low-income individuals predicted less accurately). Intersectional analysis highlighted a specific bias against non-white women.
• Individual vs. Group Prediction: While LLMs demonstrated strong capability in replicating macro-level distributions (high JSS), they struggled to capture the nuances of individual socio-political identities, resulting in lower individual accuracy compared to group-level trends.

Significance and Claims
The paper claims that while LLMs are powerful tools for simulating broad socio-political trends, their current deployment in underrepresented regions risks perpetuating inequities due to training data biases. The study asserts that the "U.S.-centric" nature of pre-training corpora limits the generalizability of these models, making them less effective in contexts like Chile where cultural and linguistic nuances differ.

The authors emphasize that ensuring fairness requires moving beyond generic model deployment. They argue for the necessity of context-specific strategies, including the collection of more diverse and representative training data, the incorporation of socio-demographic characteristics during fine-tuning, and the application of fairness constraints. The study concludes that addressing these disparities is essential for the ethical application of AI in social science research, ensuring that models do not systematically disadvantage specific socio-demographic groups or fail to represent underrepresented populations.