Year: 2026
Type: conference-paper
Abstract: Flagship LLMs can launder fake news into fluent, claim-preserving text that bypasses classifiers trained on humanwritten corpora. We quantify how three current flagships - GPT4o, Claude Sonnet 4.6, and Gemini 2.5 Pro - differ in this laundering and how the difference propagates to four detectors. Each of 1,000 ISOT fakes is laundered once per family; a threejudge ensemble isolates 906 claims preserved by all three. On this matched intersection, detector collapse is asymmetric: DeBERTav3-large falls from 0.985 to 0.738 F1-score on Claude, RoBERTalarge to 0.857 on Gemini, TF-IDF+LogReg 12.1 on Gemini, TFIDF+SVM only to 0.936 on GPT. Each detector has a different worst-case family.
Author: Jalal Mehdiyev
Institution: Institute of Information Technology
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: bronze
OpenAlex ID: W7203822992