Abstract
Large language models (LLMs) can produce fluent text that is increasingly difficult to distinguish from human writing. This raises concerns about the reliability of information online and creates a need for methods that can identify machine-generated content.
This project investigates approaches to detecting LLM-generated text and understanding how detection systems make their predictions. We examine different detection methods, evaluate their ability to distinguish human-written from machine-generated passages, and investigate factors that affect their performance. We also explore the interpretability of these systems, including the evidence they use to classify text and how explanations can support human judgement.
Together, these investigations aim to better understand the capabilities and limitations of automated text detection, with a focus on detection accuracy, reliability and interpretability.
Videos 1
Watch presentations, demos, and related content
Like, comment, and subscribe on YouTube to support the creator!
Gallery 1
Explore the visual story of this exhibit