Feature-Inversion Trap: New Benchmark Exposes LLM Detector Failures on Personalized Text

Researchers have introduced the first benchmark for detecting personalized machine-generated text (MGT), revealing that current detectors suffer significant performance drops when faced with LLM-generated imitations of a specific author's style. The study, published on arXiv (2510.12476v3), identifies a 'feature-inversion trap' where features that work for general MGT become misleading in personalized contexts. The benchmark, built from literary and blog texts paired with LLM-generated imitations, shows that even state-of-the-art detectors can fail. The authors propose a simple method to predict detector reliability. This work addresses the growing risk of identity impersonation as LLMs become more adept at imitating personal writing styles.