ETD

• AI-text detection
• authorship attribution
• computational linguistics
• discourse analysis
• machine learning
• text analytics

This thesis investigates AI-text detection through a cognitively and pragmatically grounded framework that aims to reconcile detection performance with interpretability. While current state-of-the-art detectors largely rely on opaque, representation-based or probability-driven methods, such approaches suffer from limited robustness, poor generalisation, and reduced human interpretability. The work proposes an alternative interpretable methodology that models discourse as the outcome of regulated cognitive processes characteristic of human writing. Building on a stylometric baseline, the thesis operationalises several cognitive-pragmatic dimensions—referring expressions and coreference, metacognition and metadiscourse, coherence and cohesion, temporal reasoning, and perplexity—as computationally traceable features. These feature families are evaluated individually and jointly on a large, adversarially robust benchmark to assess their discriminative power, complementarity, robustness, and interpretability. The results show that cognitively motivated features capture meaningful and partially orthogonal signals that remain informative under domain shifts and adversarial conditions, offering a theoretically grounded and human-interpretable contribution to AI-text detection.