Inside Generative AI: Interpreting Model Behaviour to Improve Safety

Generative AI is increasingly considered for use in settings where inaccurate or inappropriate outputs can have serious consequences, including critical infrastructure. Although we understand how large language models are built and trained, it remains difficult to explain why a model produces a particular response. 

This thesis will investigate methods for examining a model’s internal activity and relating it to observable behaviour. The student will select an openly available language model and a focused behaviour, such as expressing uncertainty, refusing unsafe requests, or handling sensitive information. Using interpretability tools, the student will explore which internal patterns are associated with that behaviour and test how reliably the findings hold across different prompts. The goal is to assess how insights into model behaviour could inform more effective safeguards, while recognising the limits of current interpretability methods.