As Retrieval-Augmented Generation (RAG) models gain traction in various domains, their integration of external knowledge sources with Large Language Models (LLMs) poses significant ethical challenges, particularly around bias and fairness. This PhD project aims to develop and test bias detection and mitigation methodologies that assess both the pre-trained LLMs and the external content retrieved during generation.

Your work will directly contribute to the PHAWM Project’s mission of participatory AI auditing, ensuring stakeholders – including domain experts, regulators, and end-users – can effectively audit and mitigate harms within AI systems. You will design indirect prompting methods and role-specific simulations to reveal hidden biases and deliver scalable solutions that align with global AI governance standards like the EU AI Act.

Objectives

• develop scalable tools to measure biases introduced by both LLMs and retrieved external content.
• design indirect prompting techniques that elicit implicit biases without triggering evasive model behaviours.
• simulate real-world contexts (e.g., fact-checker or policy analyst roles) to assess systemic biases in practical applications.
• optimise retrieval mechanisms to reduce biases and promote fairness by analysing the influence of external knowledge sources.
• collaborate with stakeholders to align findings with participatory harm auditing methodologies and develop real-world mitigation strategies.

This project provides a unique opportunity to collaborate with a large consortium of academic and industry partners to design real-world solutions that ensure AI systems remain ethical, fair, and trustworthy.