The AI Detector’s Toolbox : building an infrastructure for AI-generated text detection

The advent of Large Language Models (LLMs) has enabled the generation of text with remarkable fluency and coherence, leading to their widespread adoption across diverse domains. However, the increasing presence of AI-generated text (AIGT) on online platforms, in research publications, and across the digital environment more broadly has raised pressing concerns regarding misinformation, plagiarism, and overall quality standards. In particular, recent studies have reported a surge in low-quality scientific papers, the spread of misinformation across news outlets, and technical risks such as model collapse, a phenomenon in which recursive training on synthetic data degrades model performance. Consequently, the reliable detection of AIGT has emerged as a crucial research challenge in Natural Language Processing (NLP). This thesis contributes to this line of research by introducing the AI Detector’s Toolbox, a framework for advancing AIGT detection through four key components. First, I propose a new multilingual dataset that expands coverage to underrepresented languages, particularly German, and unifies existing resources into a standardized hub publicly available on HuggingFace. Second, I present LuminarSeq, a sequence-based AIGT detector that leverages a novel Intermediate Likelihood (IL) feature space and a CNN–LSTM architecture to improve upon traditional document-level detection approaches. Despite LuminarSeq’s small footprint, training studies on multiple benchmarks demonstrate competitive performance compared to other baseline models, while also providing new insights into the proposed IL feature space. Third, I develop an easy-to-use web portal that enables researchers to deploy, test, and share AIGT models seamlessly. Finally, I integrate LuminarSeq into the Docker Unified UIMA Interface (DUUI), ensuring interoperability, scalability, and reusability across research workflows. By doing so, this thesis advances the state of AIGT detection while providing standardized resources and tools to support both practical applications and theoretical insights in future research.