Style mimicry breaks AI detectors, creating a plagiarism blind spot
epoch ai found that detectors like gptzero and originality.ai fail to spot up to 29% of ai-written scientific texts when models mimic an author's style, rendering these tools unreliable for academia.
epoch ai tested pangram, gptzero, and originality.ai against a corpus of 495 human-written and 297 ai-generated passages using frontier models claude opus 4.8, gpt-5.5, and gemini 3.1 pro. while detectors caught plain ai generation with over 99% accuracy, the false-negative rate jumped to 13% on average when ai mimicked specific writing styles, peaking at 48% failure for pangram on gemini-generated scientific texts. this indicates that stylistic mimicry effectively bypasses current detection methods, leaving academia and publishers vulnerable to undetectable ai plagiarism.