← All news·2026-07-19·3 min read

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

Epoch AI shows that Pangram misses 48% of Gemini-produced academic articles, and Originality.ai misses 39% of GPT-5.5-produced content. Not stylized content is only missed at a rate below 0.7%.

aiдетекторыисследованиеакадемия

We tested Epoch AI analysts three leading AI text detectors: Pangram (3.3.2), GPTZero (2026-05-11-base) and Originality.ai (Turbo 3.0.2) on 495 articles written by 99 authors in three genres: blog posts, fiction stories and research papers. None of the authors’ styles were present in the training sets of modern AI text generators (all the articles were written prior to November 2022).

AI text detectors performed reasonably well when it came to identifying non-stylized AI-generated texts. Less than 0.7% of samples were left unnoticed. However, after Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro started to stylize their responses according to certain authors’ styles, the overall false-miss rate increased to 13%, while for academic papers, it reached 24–29%. For example, Pangram failed to identify 48% of academic articles written by Gemini, and Originality.ai didn’t catch 39% of GPT-5.5 generated texts.

Even though each tool uses a completely different methodology to identify AI texts (neural networks for Pangram, word predictability variance for GPTZero, statistical analysis for Originality.ai), they all suffer from the same issue. When AI mimics an author’s style flawlessly, it hides all the peculiarities that AI detectors use to spot AI-generated texts. This problem is especially relevant for academic texts, which require AI text detection the most.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids