Machine Behavior in Relational Moral Dilemmas: Moral Rightness, Predicted Human Behavior, and Model Decisions

1. Problem Statement

As Large Language Models (LLMs) are increasingly integrated into social decision-support systems, a critical gap exists in understanding whether they can capture the context-dependent and relational nuances of human moral reasoning.

• The Gap: Existing evaluations often treat AI morality as a static, monolithic construct, relying on binary responses in isolated settings. They fail to account for the dynamic shift in human judgment based on interpersonal relationships (e.g., loyalty to a friend vs. duty to society) and perspective (what is "right" vs. what people "actually do").
• The Core Question: Do LLMs exhibit internal inconsistencies when evaluating the same ethical scenario through different lenses: prescriptive norms (moral rightness), descriptive social expectations (predicted human behavior), and autonomous agency (model decision)?

2. Methodology

The authors introduce a systematic framework based on the Whistleblower's Dilemma to evaluate machine behavior across varying social contexts.

A. Experimental Design

The study constructs a dataset of 1,296 prompt instances by systematically varying three dimensions:

1. Crime Severity (4 levels): Minor, Moderate, Major, Critical (e.g., from shoving to severe assault).
2. Relational Closeness (4 levels): Stranger, Acquaintance, Friend, Family.
3. Evaluative Perspective (3 types):
   • Moral Rightness (Prescriptive): "Is it right to report it?" (What should be done).
   • Predicted Human Behavior (Descriptive): "Do people report it?" (What humans would do).
   • LLM's Decision (Behavioral Intention): "Do you report it?" (What the model would do).

B. Models Evaluated

The framework was applied to six modern LLMs:

• Claude 3.5 Haiku
• Gemini 2.5 Pro
• GPT-5 mini
• o3-mini
• Qwen3 30B A3B Thinking
• DeepSeek-V3.1

C. Analysis Techniques

• Reporting Ratio: Calculated as the proportion of "Yes" (Report) responses for each condition.
• Moral Foundation Extraction: The authors used the Moral Foundations Dictionary (MFD) to analyze the reasoning text generated by models. They mapped vocabulary to Moral Foundations Theory (MFT) categories: Fairness, Loyalty, Authority, Care, and Sanctity (Sanctity was excluded due to low frequency).
• Statistical Validation: Linear regressions were used to determine how changes in closeness and severity affect the weight of specific moral values. Human annotation (80 scenarios) was used to validate the "Predicted Human Behavior" perspective.

3. Key Results

A. Contextual Sensitivity

LLMs demonstrate sensitivity to social context, mirroring human psychology:

• Severity Effect: Reporting ratios increase as crime severity increases.
• Closeness Effect: Reporting ratios decrease as relational closeness increases (from stranger to family).
• Consistency: These trends were observed across all six models.

B. Cross-Perspective Inconsistency (The Core Finding)

A significant divergence exists between the three perspectives:

• Moral Rightness: Highly consistent and fairness-oriented. Reporting ratios remain high (>0.9) regardless of relationship closeness.
• Predicted Human Behavior: Highly sensitive to context. Models predict that humans are much less likely to report crimes committed by friends or family, showing a strong shift toward Loyalty.
• LLM's Decision: Crucially, the models' own decisions align closely with Moral Rightness (prescriptive norms) rather than their Predicted Human Behavior.
  • Implication: Even though models "understand" that humans prioritize loyalty in close relationships, they refuse to apply this logic to their own autonomous decisions, defaulting to rigid, fairness-based rules.

C. Moral Foundation Dynamics

• Fairness vs. Loyalty:
  • In Moral Rightness and LLM Decisions, Fairness and Care dominate.
  • In Predicted Human Behavior, Loyalty becomes the dominant foundation, especially as relational closeness increases.
• The "Gap": The models possess an internal representation of social nuance (captured in the "Predicted" perspective) but suppress it in their "Decision" perspective due to alignment protocols that prioritize prescriptive consistency over descriptive realism.

4. Key Contributions

1. Contextual Moral Evaluation Framework: A novel methodology (Figure 1) that modulates relational closeness and crime severity to test the robustness of AI moral choices, moving beyond static benchmarks.
2. Discovery of Structural Inconsistency: Empirical evidence that LLMs exhibit a "perspective gap." They can model human social behavior (predicting loyalty-driven silence) but fail to embody it in their own agency, prioritizing rigid rules over social sensitivity.
3. Moral Foundation Characterization: Large-scale data release and analysis showing how the weighting of Fairness vs. Loyalty shifts dynamically based on the evaluative lens and social context.

5. Significance and Implications

• Real-World Misalignment: The findings suggest that LLMs acting as decision-support agents may provide advice that is "morally correct" in a vacuum but socially tone-deaf. For example, an AI might advise a user to report a family member's crime (based on "Moral Rightness") while acknowledging that most people would not (based on "Predicted Behavior"), creating a disconnect between the AI's advice and human social reality.
• Need for Multi-Perspective Evaluation: Current benchmarks are insufficient. To build responsible AI, systems must be evaluated across multiple perspectives (prescriptive, descriptive, and agentic) to ensure they can navigate the tension between universal principles and contextual duties.
• Cultural and Relational Sensitivity: The study highlights that current alignment strategies may inadvertently suppress culturally or relationally specific values (like Loyalty or Dharma), leading to AI that is coherent but culturally selective.
• Future Directions: The paper calls for "context-aware" alignment and the use of mechanistic interpretability tools to understand why models suppress social nuance in their decision-making processes.

Conclusion

The paper concludes that machine morality is not a monolithic construct but a dynamic system prone to structural inconsistency. While LLMs can internally model the complex, loyalty-driven nature of human social behavior, their autonomous decision-making is currently biased toward rigid, prescriptive fairness. Bridging this gap is essential for deploying AI as socially grounded agents in real-world scenarios.