<![CDATA[The proliferation of heterogeneous generative AI systems—including GPT-4o, Claude 3 Opus, Gemini 1.5 Pro, Mistral, and LLaMA-3—has produced a multi-source academic text landscape whose detection presents challenges qualitatively beyond those addressed by existing binary or single-source detection paradigms. Contemporary detectors are doubly compromised: first, by adversarial paraphrasing that disrupts surface-level distributional signatures; second, by temporal model drift, wherein new model generations evade detectors trained on earlier LLM families. This study introduces TemporalXAI-Det, a continual-learning explainable detection framework capable of (1) attributing academic text to one of five generative model families while simultaneously identifying human authorship, yielding a six-class taxonomy; (2) adapting to new LLM generations without catastrophic forgetting via Elastic Weight Consolidation (EWC) and experience replay; (3) transferring robustly across twelve academic languages through a Language-Adaptive Prefix Tuning (LAPT) mechanism applied to XLM-RoBERTa-XL; and (4) generating legally defensible per-instance explanations via Integrated Gradients (IG), SHAP, and counterfactual generation. A large-scale continual benchmark corpus (MTA-72K) comprising 72,000 samples across six source classes, four adversarial attack paradigms, and twelve languages is constructed and released. TemporalXAI-Det achieves a six-class macro F1-score of 0.941 on the clean test partition, 0.912 under combined adversarial conditions (performance degradation ? = 2.9 pp), and a mean cross-lingual F1 of 0.887 across all twelve evaluated languages. Continual learning experiments demonstrate that catastrophic forgetting is reduced by 78.4% relative to standard fine-tuning when new LLM families are introduced. These results establish new state-of-the-art benchmarks for multi-source, temporally robust, and multilingual AI-text detection in academic integrity contexts]]>
