A Misaligned Compass: Evaluating the Conventional Bias and Moral Divergence of Large Language Models

Abstract

Recent advances in large language models (LLMs) have enabled impressive performance across reasoning tasks, yet their capacity for autonomous moral judgment remains underexplored. Prior research has not systematically benchmarked LLMs against expert standards of postconventional moral reasoning. This study aims to evaluate the moral reasoning capabilities of twelve state-of-the-art LLMs by comparing their responses to complex socio-moral dilemmas with those of a moral psychology expert, using Kohlberg’s framework as a referential standard. Models were prompted with four dilemmas (passive euthanasia, incurable diseases decisions, population control, and capital punishment) across three zero-shot tasks: moral recognition, option selection, and justification. Responses were coded using Kohlberg’s levels of moral development and statistically compared to expert judgments. Quantitative metrics included weighted kappa for concordance and logistic regression for explanatory depth. Qualitative analysis examined the structure and coherence of LLM justifications. LLMs predominantly exhibited conventional reasoning (47.7%), diverging sharply from the expert’s consistently postconventional judgment. Concordance was weak to negative, with kappa coefficients ranging from –0.333 to 0.286, indicating systemic misalignment. Binary logistic regression analysis showed that model classification significantly predicted the capacity for moral articulation; the odds ratio was 7.00 (p = 0.005), accounting for 23.4% of the variance according to Nagelkerke’s R² = 0.234. The Kruskal–Wallis test yielded H = 20.34 with p < .05, confirming significant differences in moral reasoning levels between the models and the expert. Finally, a qualitative analysis revealed that although 84.3% of justifications aligned with the selected options, they often lacked principled depth, replicating normative patterns and rejecting postconventional stances through lower-order reasoning. In conclusion, contemporary LLMs reproduce social conventions rather than engaging in autonomous ethical deliberation. The observed limitations have substantive implications for the ethical design and deployment of AI systems in morally sensitive domains. To address this, the study proposes six complementary strategies for ethical alignment, aiming to foster principled moral reasoning in AI systems: staged fine-tuning, symbolic ontology integration, expert supervision, reinforcement learning with philosophical rewards, adversarial training, and continual calibration.