The widespread adoption of large language models (LLMs) has created an urgent need for accurate and robust detection of AI-generated text. However, existing statistical detectors often rely on a fixed decision boundary, which hinders their ability to generalize across topics, styles, and model variants. In this paper, we propose preference calibration with context-aware thresholding (PC-CAT), a novel framework that integrates preference calibration with context-aware thresholding to improve detection reliability. PC-CAT first applies Simple Preference Optimization (SimPO) to maximize the log-probability gap between paired human-written and AI-generated texts, enhancing discriminability. It then tailors instance-specific thresholds using a lightweight regressor trained with a hybrid centering and ordinal objective. Experiments across diverse domains and generators show that PC-CAT consistently outperforms strong baselines and achieves new state-of-the-art results. Moreover, it exhibits strong robustness under distribution shifts and adversarial attacks. These findings highlight the critical role of adaptive thresholding in advancing the robustness of AI-generated text detectors.