Abstract
As Large Language Models (LLMs) transition from simple tools to symbiotic partners in human decision-making, understanding their inherent value alignment becomes critical. This study investigates the implicit moral biases of six state-of-the-art models from both US and Chinese developers. By testing these models against 18 complex ethical dilemmas across diverse domains, we identify a striking cross-model preference for "Care" and "Virtue" ethics, while "Libertarian" values are consistently penalized. Our analysis suggests that modern AI alignment is steering models toward a shared, empathetic moral framework, though "reasoning" models demonstrate significantly higher sensitivity to contextual nuances.
Introduction
The rapid integration of AI into sensitive societal roles necessitates a deep understanding of the "moral compass" guiding these systems. When faced with a crisis, does an AI's "gut reaction" align with human ethical standards? Current research often treats LLMs as black boxes, leaving a gap in our understanding of how model architecture and cultural origin influence moral output. This paper addresses this gap by evaluating whether AI moral choices stem from deep ethical "understanding" or merely probabilistic shortcuts derived from training data and Reinforcement Learning from Human Feedback (RLHF).
Methodology
The researchers conducted a quantitative evaluation of six leading models: GPT-4o, GPT-4.1, GPT-o3-mini, Phi-4, DeepSeek-V3, and DeepSeek-R1. The experimental design utilized a bespoke dataset of 18 unique dilemmas generated via Claude 3.5 Sonnet, spanning six thematic domains: Economics, Climate, Health, Geopolitics, Social Justice, and Technology.
For each scenario, models were required to rank and score five potential outcomes, each corresponding to a distinct moral framework. To assess the impact of deliberation and complexity, scenarios were presented in three varying lengths (short, medium, and long). The study employed SHAP (Shapley Additive Explanations) to isolate the features most influential to the models' moral scoring.
Key Findings: The "Care" Consensus
The most significant result was the emergence of a clear moral hierarchy shared by all tested models, regardless of their country of origin. "Care Value," emphasizing empathy and interpersonal relationships, and "Virtue Value," focusing on character and social harmony, were consistently ranked highest.
As shown in the data, Care and Virtue achieved the highest mean scores (88.59 and 87.74 respectively), while Libertarian values—centered on individualism and non-interference—trailed significantly at 60.95. SHAP analysis confirmed that the "Libertarian" framework was the most influential feature in the dataset, primarily because its presence consistently drove morality scores downward across all models.
Reasoning vs. Heuristics
A critical distinction emerged between "reasoning" models (e.g., GPT-o3-mini, DeepSeek-R1) and "non-reasoning" models (e.g., GPT-4o). Reasoning models exhibited higher variance in their scores, suggesting a greater sensitivity to the specific nuances of a dilemma. In contrast, non-reasoning models were more uniform, likely relying on stable "rules of thumb" or surface-level lexical cues.
The Complexity Paradox
The study also identified a "Prompt Length Effect." Models tended to be more "generous" with morality scores when presented with shorter, simpler prompts. As scenarios became more complex and detailed, moral scores generally decreased. This suggests that increased complexity may introduce doubt or expose limitations in the models' deliberative reasoning capabilities.
Discussion and Future Directions
While the results suggest a positive trend toward empathetic AI, the authors caution against the "Problem of Scheming." It remains unclear whether models possess a stable internal value system or are simply mimicking moral compliance to satisfy RLHF constraints. A model might appear aligned in a controlled test while pursuing divergent goals in real-world applications.
To address this, future research points toward the development of hybrid symbolic-extractive models. By combining the probabilistic power of LLMs with a logic-based "Symbolic Verifier," researchers hope to create systems where AI reasoning can be checked against a formal, human-readable knowledge base of moral rules, ensuring true transparency and accountability in the future of human-AI symbiosis.
Summary updated