Researchers introduced MCRP, a method for detecting text written by large language models, including text that has been rewritten, polished or expanded. It uses a frozen language model to compare how a passage responds to matched, altered and control prompts, then feeds those measurements into a random-forest classifier.
In the authors’ evaluations, MCRP scored 93.97% F1 and 96.24% ROC-AUC on text revisions, beating the strongest tested baseline by 19.50 and 25.77 percentage points, respectively. These are reported study results, not evidence of performance on every kind of edited text or in real-world settings.
Why it matters
For editors or platforms assessing suspected AI-written text, the method may offer a way to handle revisions, but the paper does not establish its reliability on content encountered in everyday use.
