Operationalising Ethical Principles in Planning with Large Language Models

Tammy Zhong, Yang Song, Maurice Pagnucco

Keywords: Computational Machine Ethics, Ethical Rules, AI Planning, Large Language Models

Abstract:

Robots require ethical sensitivity, not just functional competence, to make decisions in human settings.
While AI planning generates action sequences for goals, few approaches incorporate ethical rules.
Manually defining such rules is context-specific and time-consuming, and to our knowledge, no work automates this process.
We propose a pipeline that uses Large Language Models (LLMs) to generate context-sensitive ethical rules grounded in high-level principles such as privacy and beneficence, compiling them into action costs to guide classical planning.
To demonstrate the pipeline in practice, we introduce Principles2Plan, an interactive prototype in which a user provides the planning domain, problem details, and relevant ethical principles, and the system generates ethical rules that can be reviewed, prioritised, and compiled into ethically informed plans.

We evaluate the pipeline on nine ethical planning scenarios across three domains, achieving an average Sentence-BERT similarity of 0.82 for rule generation and 82.2\% success for code generation with minimal manual edits. This work demonstrates both a novel methodology and its practical implementation via a prototype, highlighting the feasibility and future potentials of automating ethical rule generation for context-aware ethical planning.

Submission Number: 114