MMIDR: multi-scale mutual information for AI detection via rewriting

with an unwavering adherence to stylistic uniformity throughout the entirety of their generative machinations.This stands in stark contrast to the mercurial nature of human expression, which is subject to the vagaries of individual idiosyncrasy and the capricious influence of myriad exogenous factors., i.e., the model perceives the best practice.This best practice is favoured by the model during the rewriting process.Predicated upon the aforementioned empirical observations and theoretical exegesis, this scholarly exposition proposes to elucidate a groundbreaking methodological paradigm for detection and analysis: Multi scale Mutual Information for AI Detection via Rewriting (MMIDR).This method achieves effective recognition of AI generated content by performing the multi-scale rewriting tasks on the test texts and analyzing the information changes of samples before and after rewriting.The MMIDR approach is unique relative to some past means in that it does not introduce additional training overheads and costs., but fully utilizes the black box characteristics of LLM.By analysing the generative patterns and rules of the LLM itself, the intrinsic patterns of the text to be tested are analysed at multiple scales.