Leveraging Personas to Study and Control Ideological Biases in Large Language Models

Large language models (LLMs) are increasingly embedded in sensitive domains such as information access, content moderation, and computational social science, where fairness and neutrality are critical. However, these systems consistently exhibit political and ideological biases that risk undermining trust, fairness, and resilience against manipulation. Existing research largely treats such biases as static properties of model weights, yet recent evidence suggests that prompting LLMs with personas exposes a dynamic, controllable layer of ideological expression. This dissertation investigates the mechanisms and implications of persona-driven bias along three dimensions: (i) how personas influence the expression of political and cultural biases; (ii) what strategies can mitigate or guide persona-driven ideological shifts; and (iii) whether such dynamics generalise across broader ideological domains. Our initial findings show that personas hold the potential to substantially alter ideological positioning, with larger models demonstrating broader dispersion and stronger responsiveness to explicit cues. Moreover, thematic persona attributes (e.g., business, history, politics) consistently trigger predictable ideological shifts, revealing latent pathways of bias expression. Building on these insights, forthcoming work will explore bias mitigation strategies—such as direct preference optimization (DPO)—and tests their effectiveness in reducing political and cultural distortions while preserving task performance. By uncovering both the risks and opportunities of persona-driven ideological malleability, this research contributes to the development of more trustworthy, fair, and context-aware generative AI systems.