Can AI Detectors Tell If Text Was Paraphrased? Accuracy Benchmarks and Student Risks