A rigorous study by Epoch AI has revealed a critical blind spot in the AI text detectors increasingly relied upon by universities, publishers, and employers: when AI models are prompted to imitate a specific author's writing style, detectors miss between 10 and 18 percent of the resulting passages, with scientific writing proving especially vulnerable at up to 26 percent undetected.

The study, published July 15, 2026, tested three of the most prominent commercial detectors — Pangram, GPTZero, and Originality.ai — against 297 AI-generated passages produced by Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. Each model generated text under two conditions: a basic prompt and a style-imitation prompt where the model was given five samples of a real author's work and asked to write in the same style. A corpus of 495 genuine human-written passages from 99 authors across blogging, fiction, and scientific writing served as the baseline.

On basic, unadorned AI prompts, the detectors performed near perfectly — false-negative rates were at most 0.7 percent across all three tools. Human text was also handled well: Pangram and GPTZero correctly identified all 495 human passages as human, while Originality.ai falsely flagged 19 passages (3.8 percent).

But the picture changed sharply under style imitation. Originality.ai missed 18 percent of style-mimicked passages, GPTZero missed 11 percent, and Pangram missed 10 percent. The evasion was most dramatic in scientific writing, where Originality.ai failed to flag 29 percent of AI-generated scientific passages, and Gemini 3.1 Pro's scientific output evaded Pangram 48 percent of the time.

The technique required nothing more sophisticated than a simple prompt — making an AI model imitate a writer's style is trivially easy, while detecting the result is a hard statistical problem that grows harder as frontier models improve. Epoch AI's Jaeho Lee, who authored the study, noted that "Mixed" verdicts from Pangram and GPTZero were conservatively counted as non-detections, meaning the true miss rate could be even higher depending on how institutions interpret ambiguous results.

The findings carry serious implications for academic integrity, publishing, and hiring. Universities deploy these detectors to police plagiarism, journals screen submissions with them, and employers evaluate written assessments — often treating detector results as decisive evidence. A tool that misses roughly one in five AI passages when a simple style-imitation prompt is applied is not a reliable basis for consequential decisions.

Sources