AI text detectors fail when AI mimics a writer’s style

Epoch AI analysts put three top AI text detection tools — Pangram (version 3.3.2), GPTZero (2026-05-11-base), and Originality.ai (version Turbo 3.0.2) — to the test with 495 articles from 99 different authors across three categories: blogs, fiction, and research papers. All the texts were produced before November 2022, so none of the authors’ styles were included in the training data for modern AI text generators.

The tools did pretty well at detecting non-stylized AI-generated text, with only less than 0.7% of samples going undetected. But once Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro started stylizing their output based on the styles of particular authors, the average false-miss rate rose to 13%, and for academic content, it jumped to 24–29%. In the worst case, Pangram missed 48% of academic articles produced by Gemini, and Originality.ai missed 39% of GPT-5.5 outputs.

Despite their completely different approaches to AI text detection — neural nets for Pangram, word predictability variance for GPTZero, statistical analysis for Originality.ai — all three tools demonstrated this same flaw. Imitating the author’s style perfectly camouflages the features these tools look for in AI-generated texts. The most affected genre is academic writing, where we need AI text detection most of all.