Why AI-Generated Text Detection Fails: Evidence from Explainable AI Beyond Benchmark Accuracy