Sentence-level AI detectors are essential for detecting AI misuse in human-AI collaborative writing. Yet existing approaches still treat all authors identically, ignoring individual writing styles. This is a critical oversight, given that the boundary between human and AI text is inherently author-specific. To address this limitation, we propose PHAS-ID (Personalized History-Aware Sentence-level AI Detector), an approach that introduces personalization to sentence-level detection. By matching target sentences against the author’s history of labeled documents, PHAS-ID tailors its predictions based on prior writing behaviors. Furthermore, current synthetic datasets for this task fail to capture realistic collaborative patterns, which we address by introducing GhostFill, a style-aware synthetic dataset generation pipeline. Our experiments demonstrate that PHAS-ID outperforms the previous state-of-the-art (SOTA) across seen-author, unseen-author, cross-domain, and cross-generator settings. Ablation studies confirm that these improvements stem from genuine author-specific personalization rather than the mere addition of context. Our source code and dataset are available at our GitHub repository.

Unlike document-level AI-generated text (AIGT) detection, sentence-level AIGT detection remains underexplored, despite its importance for addressing collaborative writing scenarios where humans modify AIGT suggestions on a sentence-by-sentence basis. Prior sentence-level detectors often neglect the valuable context surrounding the target sentence, which may contain crucial linguistic artifacts that indicate a potential change in authorship. We propose GL-CLiC, a novel technique that leverages both Global and Local signals of Coherence and Lexical Complexity, which we operationalize through discourse analysis and CEFR-based vocabulary sophistication. GL-CLiC models local coherence and lexical complexity by examining a sentence’s relationship with its neighbors or peers, complemented with its document-wide analysis. Our experimental results show that GL-CLiC achieves superior performance and better generalization across domains compared to existing methods.