AMULED: Addressing Moral Uncertainty using Large language models for Ethical Decision-making.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Ethics in AI

Background:

• Reinforcement learning (RL) often struggles with moral uncertainty due to reliance on single ethical frameworks or handcrafted rewards.
• Existing methods lack scalability and fail to capture the complexity of moral pluralism in AI decision-making.

Purpose of the Study:

• To introduce AMULED, a task-agnostic ethical layer for RL agents.
• To refine pre-trained RL agents using multi-perspective moral feedback from large language models (LLMs).
• To operationalize moral pluralism and resolve conflicting ethical signals in AI.

Main Methods:

• Fine-tuning RL models with LLM-generated feedback, incorporating five moral clusters: consequentialist, deontological, virtue, care, and social justice.
• Aggregating ethical beliefs using Belief Jensen-Shannon Divergence and Dempster-Shafer Theory (BJSD-DST) to generate shaping rewards.
• Utilizing KL-regularization to maintain policy stability during ethical refinement.

Main Results:

• AMULED significantly improved ethical actions in simulated environments (e.g., attending to more crying babies, rescuing more targets) while minimally impacting task performance.
• The BJSD-DST aggregation method outperformed standard approaches in handling conflicting moral signals.
• Evaluations across different LLM backbones and environments demonstrated the framework's robustness.

Conclusions:

• AMULED provides a scalable, LLM-driven solution for integrating moral pluralism into RL decision-making.
• LLM-based belief aggregation offers a practical alternative to human supervision and handcrafted rewards for ethical AI.
• Future work should address performance variations in complex environments and LLM reasoning quality.