Can We Trust AI Agents? A Case Study of an LLM-Based Multi-Agent System for Ethical AI

TL;DR

This paper tackles the challenge of operationalizing AI ethics within LLM-driven software development by examining trustworthiness. It adapts Design Science Research to identify four trustworthiness-enhancing techniques—multi-agent collaboration, specialized roles, structured communication, and multiple rounds of debate—and builds an LLM-based multi-agent system (LLM-MAS). The prototype is evaluated on three real-world AI incidents using thematic analysis, hierarchical clustering, a baseline comparison, and runnable code, yielding around 2,000 lines of ethically informed output per case versus ~80 lines for baseline prompts. Findings show the approach can generate extensive code and documentation addressing bias, transparency, GDPR, and EU AI Act compliance, but highlight practical integration and dependency-management challenges that limit immediate practitioner adoption. Together, these results offer a concrete path to operationalize AI ethics in AI4SE while signaling areas for further refinement and open-sourcing to enhance practicality and reproducibility.

Abstract

AI-based systems, including Large Language Models (LLM), impact millions by supporting diverse tasks but face issues like misinformation, bias, and misuse. AI ethics is crucial as new technologies and concerns emerge, but objective, practical guidance remains debated. This study examines the use of LLMs for AI ethics in practice, assessing how LLM trustworthiness-enhancing techniques affect software development in this context. Using the Design Science Research (DSR) method, we identify techniques for LLM trustworthiness: multi-agents, distinct roles, structured communication, and multiple rounds of debate. We design a multi-agent prototype LLM-MAS, where agents engage in structured discussions on real-world AI ethics issues from the AI Incident Database. We evaluate the prototype across three case scenarios using thematic analysis, hierarchical clustering, comparative (baseline) studies, and running source code. The system generates approximately 2,000 lines of code per case, compared to only 80 lines in baseline trials. Discussions reveal terms like bias detection, transparency, accountability, user consent, GDPR compliance, fairness evaluation, and EU AI Act compliance, showing this prototype ability to generate extensive source code and documentation addressing often overlooked AI ethics issues. However, practical challenges in source code integration and dependency management may limit its use by practitioners.

Figures (5)

• Figure 1: Prototype overview
• Figure 2: Evaluation overview
• Figure 3: Hierarchical clustering dendogram - PD1
• Figure 4: Hierarchical clustering dendogram - PD2
• Figure 5: Hierarchical clustering dendogram - PD3