AI text detectors struggle when language models mimic an author's style

Research by Epoch AI indicates that detection tools like Pangram, GPTZero, and Originality.ai frequently fail to identify content when large language models are prompted to mimic a specific author's writing style. While these tools missed up to 18 percent of general AI-generated text, the error rate reached 48 percent for scientific writing. This suggests that as AI becomes more adept at stylistic imitation, current detection methods are becoming increasingly unreliable for verifying the human authorship of academic and professional documents.