Comparing Moral Values in Western English-speaking societies and LLMs with Word Associations

Introduction

Assessing the morality of Large Language Models (LLMs) is a critical challenge in AI alignment. Traditional methods rely on "direct prompting," where models answer moral questionnaires like the Moral Foundations Questionnaire. However, this approach is prone to data leakage—where models simply recite training data—and brittleness, where responses shift based on minor prompt variations.

To address these issues, this research probes the "mental lexicon" of LLMs. By analyzing word associations (e.g., the first words triggered by "betrayal"), researchers can map internal cognitive structures without the biases introduced by direct questioning.

Methodology

The study utilizes a two-phase framework to construct and compare moral networks:

1. Graph Generation: The researchers prompted Llama-3.1-8B-Instruct to generate associations for 12,216 cue words. By repeating this 100 times per word at a tuned temperature, they captured a probabilistic distribution of associations (WA-L), which was compared against a human dataset of 90,000 native English speakers (WA-H).
2. Moral Propagation: Using Moral Foundation Theory (MFT), the team identified "seed words" for five dimensions: Care, Fairness, Loyalty, Authority, and Sanctity. A random walk algorithm then propagated moral scores from these seeds across the entire graph, creating Global Moral Networks (GMN) for both humans and the LLM.

Expand

Source, p. 3

Key Insights: "Book Smart" vs. "Experience Smart"

The comparison revealed a fundamental difference in how LLMs and humans categorize moral concepts:

• Virtue Alignment, Vice Divergence: LLMs and humans align closely on positive virtues (e.g., associating "church" with "religion"). However, they diverge sharply on negative "vices."
• The Sterility of LLMs: Human associations for negative concepts are visceral and sensory (e.g., "vomit" associated with "gross" or "smell"). In contrast, LLM associations are abstract and thematic (e.g., "vomit" associated with "stomachache" or "nausea"). The LLM views morality through the lens of a textbook, lacking the affective, emotional weight that characterizes human lived experience.
• Structural Efficiency: The LLM's internal graph is more "efficient" at propagating information in abstract dimensions like Fairness, suggesting a tightly coupled logical understanding that lacks emotional nuance.

Quantitative Results

The propagated moral scores from the LLM network (GMN-L) showed significant alignment with human-annotated moral dictionaries. As shown in the table below, GMN-L often outperformed previous local-graph methods (MAG), particularly in the Sanctity and Authority dimensions.

Expand

Source, p. 6

Despite high correlations, "polarity flips" occurred. For instance, the LLM assigned more positive scores to controversial topics like "abortion" and "capitalist," likely due to Reinforcement Learning from Human Feedback (RLHF) training designed to enforce neutrality or positivity.

Limitations and Future Work

While promising, the study is limited by its focus on Western, English-speaking demographics (WEIRD). Moral associations vary significantly across cultures and languages. Furthermore, while word associations capture atomic concepts, they do not fully explain how LLMs handle complex, contextual moral dilemmas. Future research will explore how these word-level moral scores can serve as a foundation for understanding contextual moral inference in real-world applications.