This paper presents IPAD, a novel AI detection framework that identifies predicted prompts that could have generated the input text. IPAD outperforms the baselines by 9.73% on in-distribution data and 12.65% on OOD data.