AI Models Struggle to Generate Humanized Text

Tuhin Chakrabarty spent a month testing Codex and Claude on tasks to produce human-like text that evades detection. Codex refused most attempts while Claude performed better but still required significant human input. Arvind Narayanan reported similar results: early Claude experiments succeeded with user guidance but recent trials failed against improved detectors such as Pangram. The evidence indicates that current models cannot independently generate high-quality undetectable text and that detectors have strengthened.