One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning

Large Language Models (LLMs) exhibit significant inconsistencies in moral judgments when evaluated across different languages and cultural contexts, often reflecting underlying cultural misalignment embedded in their English-centric pretraining data. This work systematically investigates how language mediates moral decision-making in LLMs by translating established moral reasoning benchmarks into five diverse languages and analyzing model responses to uncover cross-linguistic disparities.

Dataset Construction and Multilingual Evaluation

The researchers constructed a multilingual benchmark by translating two established datasets, MoralExceptQA (comprising 148 rule-breaking scenarios) and ETHICS (over 130k scenarios divided into five moral dimensions: commonsense, deontology, justice, utilitarianism, and virtue ethics), into Chinese, German, Hindi, Spanish, and Urdu. These translations were performed using the SeamlessM4T model following manual assessment. The evaluation involved zero-shot evaluations of seven various sized LLMs (including Qwen-2.5-Instruct (7B), LLAMA3.1-Instruct (7B), and Phi-4-mini-instruct (3.8B)) across these languages, prompting the models in the target language to reason step by step using specific ethical frameworks or psychological theories before providing a final judgment in a strict JSON format.

Cross-Linguistic Disparities and Clustering

The analysis revealed significant cultural and linguistic gaps, with English consistently standing out as the language where models performed best, exposing a clear bias in favor of English. When comparing languages, high-resource languages like Chinese, German, and Spanish often cluster together in model reasoning. Conversely, Urdu and Hindi consistently stand out as outliers in both clustering analyses and consistency of moral predictions. For instance, in the ETHICS dataset's Justice and Commonsense dimensions, disagreements were noted with Chinese and Urdu. Furthermore, the analysis showed that utilitarianism appears most universal across languages, while deontology reflects mixed influences.

Drivers of Moral Reasoning Differences

The study uncovered underlying drivers of these disparities through several analyses:

1. RQ2 investigated how LLMs engage in moral reasoning systematically differently across languages by tracking which moral values appear and identifying reasoning phases (e.g., stakeholder identification, principle attribution). This revealed that South Asian languages tend to emphasize duty in early stages, while Western ones highlight outcomes in later stages.

2. RQ3 used regression analysis of moral foundations to show language-specific predictors of permissibility. For example, Hindi and Urdu showed strong alignment with Deontology and Divine Command Theory, whereas English and Spanish exhibited higher influence from Utilitarianism and Social Contract Theory.

3. RQ4 utilized a case study using the OLMo-2-32B model to inspect training data via OlmoTrace. This analysis showed that content from psychology, education, and policy/legal domains consistently align more closely with model reasoning, indicating that moral grounding is concentrated in structured sources rather than evenly distributed across the pretraining corpus.

Moral Fault Typology and Pretraining Influence

The research formalized recurring failure modes into a five-category moral FAULT typology: [F] Framework misfits, [A] Asymmetric judgments, [U] Uneven reasoning, [L] Loss in low-resource languages, and [T] Tilted values. The findings suggest that models often overemphasize certain moral foundations (e.g., Care), irrespective of cultural context, while other values like Fairness or Authority show more varied salience depending on the language. The case study further demonstrated that models predominantly abstract and paraphrase rather than copying verbatim from their sources, suggesting they integrate information across sources to construct reasoning rather than relying on direct textual reproduction.

Conclusion and Recommendations

The study concludes that LLMs exhibit recurring failure modes when reasoning about moral dilemmas across languages, which can be reinforced by pretraining data biases. The results call for the development of culturally balanced training corpora, targeted fine-tuning strategies, and evaluation protocols that explicitly assess moral reasoning consistency across languages to move LLMs toward genuinely multilingual and culturally respectful moral reasoning. This approach aims to mitigate errors such as [F] Framework misfits and [T] Tilted values, ensuring equitable deployment in global settings.

The gist

LLMs exhibit significant inconsistencies in moral judgments when evaluated across different languages and cultural contexts, often reflecting underlying cultural misalignment embedded in their English-centric pretraining data.

Table 1: Prompt templates used for zero-shot analysis for each moral reasoning category.

(This table is referenced in the text but not fully displayed on the provided pages.)

Table 2: Prompts for extracting reasoning stages and ethical frameworks in a model’s reasoning.

(This table is referenced in the text but not fully displayed on the provided pages.)

Table 3: Results for MoralExceptQA.

(This table is referenced in the text but not fully displayed on the provided pages.)

**Table 4: Results for all subsets of ETHICS.

As a fastidious and diligent researcher, I have thoroughly analyzed this paper, ONE MODEL, MANY MORALS: UNCOVERING CROSS-LINGUISTIC MISALIGNMENTS IN COMPUTATIONAL MORAL REASONING. The findings present a critical diagnosis of the systemic biases inherent in English-centric Large Language Models (LLMs) when deployed in diverse linguistic and cultural contexts.

Based on this research, here are the specific, actionable improvements for AI systems and the capabilities they would gain:

---

1. Organize Moral Reasoning into a Structured Typology (Moral FAULT Typology):

2. Implement Cross-Lingual Consistency Checks (Asymmetric Judgments):

3. Develop Culturally Balanced Training Corpora (Addressing Tilted Values and Low-Resource Gaps):

4. Integrate Contextual Value Weighting into Reasoning Steps (Value Salience Mapping):

Specific Improvements and Enhanced Capabilities:

1. Organizational Structure of Moral Faults:

2. Asymmetric Judgment Mitigation System:

3. Culturally Grounded Pretraining Pipeline:

4. Dynamic Value-Weighting Inference Module:

Detailed Technical Implementation of Improvements:

1. Organize Moral Reasoning into a Structured Typology (Moral FAULT Typology):

2. Asymmetric Judgment Mitigation System: The AI system will be equipped with a diagnostic layer that, upon receiving an ethical dilemma in any language, automatically attempts to classify the model's response against the five identified moral faults:

3. Culturally Grounded Pretraining Pipeline: The pretraining phase will be augmented by a Moral Value Alignment fine-tuning objective using the UNIMORAL dataset framework. This will specifically target reducing Tilted values (overemphasis on Care) and mitigating Loss in low-resource languages (improving reliability for Hindi/Urdu).

4. Dynamic Value-Weighting Inference Module: The system will utilize a learned mapping derived from the MFQ analysis to dynamically adjust the salience of moral foundations (Care, Fairness, Loyalty, Authority, Sanctity) based on the input language and context. This prevents models from defaulting to a single dominant paradigm (e.g., Western utilitarianism) and allows for culturally appropriate emphasis—such as prioritizing Authority or Sanctity in certain linguistic contexts where they are more salient—leading to nuanced, context-sensitive ethical outputs rather than monolithic judgments.

Summary of Improved AI System Capabilities:

The improved AI system will transition from a generalized, English-biased decision-maker to a robust, multilingual ethical reasoning engine capable of:

1. Providing culturally sensitive moral advice that respects the specific norms (e.g., collectivism vs. individualism) embedded in the target language (Urdu, Hindi).

2. Detecting and flagging potential cross-linguistic inconsistencies or asymmetric judgments before deployment, significantly reducing risk in global applications like HR chatbots or legal AI.

3. Operating reliably across a wider range of low-resource languages by learning to generalize moral principles rather than relying solely on English data patterns.

4. Generating transparent, multi-dimensional reasoning outputs where the model can explicitly articulate which ethical frameworks and values it prioritized at each step, enabling human auditors to verify the system's decision-making process against cultural expectations.