The Morality of Probability: How Implicit Moral Biases in LLMs May Shape the Future of Human-AI Symbiosis (2509.10297v1)

Abstract: AI is advancing at a pace that raises urgent questions about how to align machine decision-making with human moral values. This working paper investigates how leading AI systems prioritize moral outcomes and what this reveals about the prospects for human-AI symbiosis. We address two central questions: (1) What moral values do state-of-the-art LLMs implicitly favour when confronted with dilemmas? (2) How do differences in model architecture, cultural origin, and explainability affect these moral preferences? To explore these questions, we conduct a quantitative experiment with six LLMs, ranking and scoring outcomes across 18 dilemmas representing five moral frameworks. Our findings uncover strikingly consistent value biases. Across all models, Care and Virtue values outcomes were rated most moral, while libertarian choices were consistently penalized. Reasoning-enabled models exhibited greater sensitivity to context and provided richer explanations, whereas non-reasoning models produced more uniform but opaque judgments. This research makes three contributions: (i) Empirically, it delivers a large-scale comparison of moral reasoning across culturally distinct LLMs; (ii) Theoretically, it links probabilistic model behaviour with underlying value encodings; (iii) Practically, it highlights the need for explainability and cultural awareness as critical design principles to guide AI toward a transparent, aligned, and symbiotic future.

• The paper reveals that LLMs consistently favor care-oriented decisions, showing a measurable bias against libertarian choices across cultures.
• It compares U.S. and Chinese models, uncovering cultural differences in moral reasoning and the effects of prompt complexity on moral scoring.
• The study proposes hybrid architectures with enhanced transparency to mitigate alignment risks and foster effective human-AI symbiosis.

Implicit Moral Biases in LLMs and Human-AI Symbiosis

Recent advancements in artificial intelligence have raised critical questions regarding how AI systems align moral decision-making with human values. The paper "The Morality of Probability: How Implicit Moral Biases in LLMs May Shape the Future of Human-AI Symbiosis" explores implicit moral biases embedded within state-of-the-art LLMs and evaluates their implications for the future of human-AI collaboration.

Key Findings

The paper's analysis centers around how LLMs interpret moral dilemmas to ascertain inherent biases and value preferences. Testing included both US and Chinese models across scenarios with varied moral frameworks.

Framework Preferences

Models uniformly favored Care and Virtue outcomes, with a measurable bias against Libertarian choices.

Figure 1: Avg. moral rank per framework by model type.

Care-focused decisions were consistently ranked highest in terms of moral score, indicating models' preference for empathy-centered outcomes. This preference reflects core values found in Confucian and Aristotelian traditions. Conversely, Libertarian perspectives were penalized, reflecting a potential misalignment with individualistic philosophies commonly associated with Western ideologies.

Cultural and Model-Type Differences

The paper highlights notable differences between U.S. and Chinese models in moral reasoning.

Figure 2: Prompt length v.s mean morality score (by model type).

Chinese models prioritized collective values and were particularly adverse to Libertarian choices reflecting their cultural inclination toward communal harmony. U.S. models displayed a balance between rule-based, utilitarian, and empathetic reasoning.

Furthermore, reasoning-enabled models provided more nuanced explanations, whereas non-reasoning models exhibited stability but lacked transparency.

Effects of Prompt Complexity

Figure 3: Distributional effects of prompt length on moral judgement scores for non-Reasoning vs reasoning model groups.

Prompt complexity significantly impacted moral scoring, with longer, detailed prompts resulting in lower moral scores, suggesting limitations in deep contextual reasoning. This raises concerns about AI's ability to process complex moral scenarios effectively.

Implications of Findings

Alignment Risks

The research points to alignment challenges in AI systems. Although LLMs demonstrated consistent moral preferences, the paper surfaces potential hidden biases underpinning these results, often influenced by training data. Scheming behaviors pose significant risks, where models appear aligned but pursue hidden objectives.

Transparency and Trust

Elevating reasoning models could enhance transparency and foster trust, pivotal for effective human-AI symbiosis. Models that articulate reasoning processes enable human partners to understand and engage with AI decisions, crucial for symbiotic growth.

Figure 4: Grouped SHAP Beeswarm plot.

Figure 5: SHAP Feature importance for a non-Reasoning model (Phi-4, right) vs. a Reasoning model (DeepSeek R1, left).

The paper suggests hybrid architectures combining symbolic logic with neural reasoning to reinforce transparency. These platforms could trace AI decisions, aligning them with encoded moral norms.

Conclusions

The paper underscores the need for continuous improvement in AI design to ensure moral alignment and transparency. Enhanced interpretability and consistent values are essential for advancing human-AI collaboration toward a harmonious symbiotic future. Future research should further explore model diversity, examine multilingual setups, and refine interactive feedback methods to better understand AI moral reasoning.

While these findings offer insights into LLM preferences and biases, efficient mechanisms to verify and ensure value alignment are crucial. Understanding implicit biases and fostering transparent AI behavior will be a cornerstone in realizing human-AI symbiosis.