Hidden Human-Like Nature of Machine-Generated Texts: Theory and Detection Enhancement

In three linesResearchers reveal that LLM-generated texts contain hidden human-like spans that complicate detection. They propose a stacked model-agnostic framework using a hard-EM procedure to iteratively filter human-like subsequences and enhance existing detectors, working also in training-free mode.

Summary generated by Claude — human-verified