"Can Software Development Adopt AI Ethics?" : Agents4EthicalSE: A Multi-Agent RAG Framework

Abstract

As Large Language Models (LLMs) based solutions such as GitHub Copilot, Amazon CodeWhisperer, Rubberduck (ChatGPT based) become widely used in software development, questions remain about how these tools can support ethical and regulatory compliance. While existing AI ethics guidelines outline high-level principles such as fairness, transparency, and accountability, developers lack tools that operationalize these during day-to-day coding tasks. This thesis introduces Agents4EthicalSE, a multi-agent Retrieval-Augmented Generation (RAG) framework that assists developers in generating ethically grounded code aligned with the EU AI Act, AI HLEG guidelines, and other relevant regulatory documents.

The framework is designed as a practical alternative to baseline code generation work-flows. In a typical setup, tools like Copilot produce outputs without referencing legal or ethical constraints. In contrast, Agents4EthicalSE consists of specialized agents—such as a RiskGuard and an AI Ethicist—that engage in structured multi-round interactions. The AI Ethi-cist uses RAG to retrieve relevant content from regulatory texts and guidelines, enabling real-time critique and refinement of code suggestions. Developers can review these outputs and it-erate with agent feedback to address ethical risks or non-compliance concerns.

We evaluated the system through a workshop-based study involving 82 AI practitioners and developers. Participants were asked to simulate development tasks using both Agents4EthicalSE and a baseline LLM. The tasks were derived from real-world AI failures documented in the AI Incident Database, reframed as project descriptions across different risk categories defined by the EU AI Act. Survey results showed that Agents4EthicalSE helped participants better identify ethically sensitive components, interpret compliance requirements, and reflect on ethical trade-offs during code generation. Feedback also pointed to areas for improvement, such as citation clarity and interface responsiveness.

These results suggest that multi-agent RAG frameworks can provide developers with more interpretable and regulation-aware coding assistance compared to standard LLMs. This work contributes a step toward integrating ethical reasoning into software engineering workflows through structured, document-informed interactions.