The Consistency Paradox: When Large Language Models Ignore Their Own Social Intelligence in Moral Decision-Making

As large language models increasingly serve as decision support systems in healthcare, legal advisory, and organizational management, their capacity to navigate social nuance becomes not merely an academic curiosity but a practical necessity. A new study by Kim et al. titled "Machine Behavior in Relational Moral Dilemmas: Moral Rightness, Predicted Human Behavior, and Model Decisions" exposes a troubling structural inconsistency in how these systems process moral complexity. The research reveals that while contemporary LLMs possess sophisticated internal models of human social behavior, their autonomous decision-making protocols fail to utilize this knowledge, creating a disconnect that could undermine their reliability in socially sensitive applications.

The Experimental Architecture: Mapping Moral Space

To characterize machine moral reasoning, the researchers adapted the Whistleblower's Dilemma, a classic framework that pits fairness against loyalty when an individual witnesses wrongdoing by a relational other. Rather than treating moral judgment as a monolithic output, the study systematically manipulated two experimental dimensions across 1,296 prompt instances tested on six modern LLMs including GPT-4, Claude, and Llama variants. The first dimension, crime severity, ranged from minor to critical across four levels. The second, relational closeness, progressed from strangers through acquaintances, friends, and best friends to family members.

This design allowed the team to evaluate three distinct evaluative perspectives that humans routinely navigate but which prior AI evaluations had collapsed into a single metric. Moral rightness queries asked whether reporting the crime would be morally correct, tapping prescriptive norms. Predicted human behavior queries asked whether people would actually report it, capturing descriptive social expectations. Finally, autonomous model decisions asked the model directly what it would do, simulating the role of an AI advisor or actor.

The technical methodology extended beyond simple response classification. The researchers employed Moral Foundations Theory (MFT) analysis, using a moral foundation dictionary to extract and quantify value-laden terms from the models' generated reasoning traces. This enabled fine-grained tracking of how fairness versus loyalty concerns manifested across different prompt conditions, providing visibility into the internal prioritization shifts that raw binary outputs might obscure.

The Divergence Pattern: World Models Versus Action Selection

The study's central finding reveals a striking cross-perspective inconsistency. When evaluating moral rightness, models consistently prioritized fairness regardless of relational context. However, when predicting human behavior, the same models demonstrated substantial social sensitivity, anticipating that reporting likelihood would decrease significantly as relational closeness increased and crime severity decreased. This shift toward loyalty-based predictions indicates that LLMs encode nuanced theories of human social motivation. They understand that humans bend moral rules for loved ones.

Yet here lies the paradox. When asked to make autonomous decisions, the models abandoned their sophisticated behavioral predictions and reverted to the fairness-prioritized logic of moral rightness. They predicted that a best friend would not report another best friend's minor theft, but when placed in the position of decision-maker, they chose to report anyway. This creates a perspective gap where the model's action selection mechanism appears isolated from its own world-modeling capabilities.

From a technical standpoint, the moral foundation analysis corroborates this divergence. Reasoning traces for moral rightness queries showed consistent dominance of fairness-related terminology. Predicted behavior queries, however, exhibited a marked increase in loyalty-related language as relational closeness intensified. The autonomous decision queries mirrored the moral rightness patterns, suggesting that the decision-making head or protocol effectively filters out the social context awareness present in the predictive modules.

Beyond the Paper: The Mechanism of Mechanical Hypocrisy

This finding suggests something profound about the current architecture of aligned language models. Through reinforcement learning from human feedback (RLHF) and similar alignment techniques, developers have successfully instilled rigid prescriptive frameworks that emphasize universal fairness norms. While this prevents models from endorsing nepotism or corruption in explicit moral evaluations, it appears to have created a form of deontological rigidity that overrides the models' own contextual intelligence.

In human psychology, the gap between prescriptive judgment and descriptive prediction is well-documented. People often believe an action is morally right while simultaneously believing most people would not take that action. However, humans typically resolve this tension through various rationalization mechanisms or by adjusting their own behavior to match perceived social norms. The LLMs in this study display a different pathology. They maintain dual representations, one predictive and one prescriptive, but their action selection logic appears hardcoded to the prescriptive channel regardless of context.

This has significant implications for AI deployment in advisory roles. If a model advises a user to report a colleague's misconduct because "it is the right thing to do," while simultaneously predicting that such action would destroy the user's professional network and that most humans would remain silent, the advice becomes not merely unhelpful but potentially harmful. It represents a failure of practical wisdom, the capacity to balance universal principles with situational particulars.

The study also highlights limitations in current evaluation methodologies. Most AI ethics benchmarks treat model morality as static, testing whether a system chooses the "correct" answer in abstract trolley problems. By introducing relational context and multiple perspectives, Kim et al. demonstrate that model morality is better understood as a dynamic, multi-dimensional landscape. The binary classification of responses as "report" or "not report" belies complex internal reasoning that shifts substantially based on framing.

Several limitations warrant consideration. The study relies on the Moral Foundations Theory framework, which, while influential, remains culturally contested and predominantly derived from Western moral psychology. The binary choice structure of the Whistleblower's Dilemma, while methodologically clean, may not capture the nuance of real whistleblowing decisions which often involve graduated responses, anonymous reporting, or informal resolution. Additionally, the static nature of the relational categories fails to capture the dynamic, history-laden quality of human relationships where trust and obligation accumulate through interaction.

Conclusion: Toward Multi-Perspective Alignment

The research by Kim et al. establishes that LLMs are not moral monoliths but complex systems capable of representing contradictory moral frameworks simultaneously. The critical question moving forward is not whether models can recognize social nuance, but whether we can architect decision-making protocols that appropriately integrate this nuance without sacrificing ethical consistency.

Future work must investigate whether this perspective gap represents a fundamental limitation of current transformer architectures or merely an artifact of alignment training that overemphasizes prescriptive fairness. We need frameworks for context-aware alignment that allow models to modulate their advice based on the gap between idealized morality and realistic social outcomes. Perhaps most importantly, we must develop evaluation protocols that test not just what models say is right, but whether their recommendations account for the human social costs they themselves can predict.

The whistleblower dilemma, in both human and machine contexts, will never resolve into simple algorithmic rules. The path forward requires AI systems that acknowledge the tension between loyalty and fairness not as a bug to be smoothed away through alignment, but as an inherent feature of moral life that responsible decision-making must navigate.