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

2026-07-19 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, short

Summary

A research team at Epoch AI found that popular AI text detectors, including Pangram (version 3.3.2), GPTZero (model 2026-05-11-base), and Originality.ai (Turbo 3.0.2), struggle significantly when language models mimic a specific author's writing style. While these tools achieve near-perfect accuracy on plain AI-generated text, with false-negative rates topping out at 0.7 percent, their performance drops considerably for style-imitated content. The study, which used frontier models like Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, revealed an average of 13 percent of such passages went undetected. Scientific writing proved to be the most challenging category, with detectors failing to flag between 24 and 29 percent of style-mimicking AI-generated content, raising concerns about their reliability in academic settings. Originality.ai also showed a 3.8 percent false-positive rate on human texts.

Key takeaway

For educators and academic integrity officers evaluating student submissions, relying solely on AI text detectors is insufficient. Your current tools, like Pangram or Originality.ai, are demonstrably vulnerable to AI models mimicking human writing styles, especially in scientific contexts, with up to 29 percent of AI-generated content slipping through. You should implement multi-faceted verification strategies beyond automated detection to ensure genuine authorship and prevent sophisticated AI misuse.

Key insights

AI text detectors struggle significantly when language models mimic an author's style, particularly in scientific content.

Principles

• Generic AI text is reliably detectable.
• Style imitation bypasses current detectors.
• Scientific writing poses unique detection challenges.

Method

Epoch AI tested three detectors against human, plain AI, and style-mimicked AI texts. Frontier models (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro) generated style-imitated content from a 495-passage corpus.

In practice

• Do not solely trust detectors for academic integrity.
• Manually verify scientific content for AI use.
• Recognize style imitation as an evasion tactic.

Topics

• AI Text Detection
• Large Language Models
• Style Imitation
• Academic Integrity
• Detector Performance
• Scientific Writing

Code references

• jaeholee-brown/ai-text-detectors

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, MLOps Engineer, AI Ethicist

Related on AIssential

• Why We Can’t Have a Reliable AI Text Detector — The inherent unreliability of AI text detectors makes them unsuitable for critical applications.
• How AI is changing language — AI's linguistic influence blurs human-machine writing distinctions and reshapes human language, challenging authenticity and originality.
• Q&A: Pangram CEO — AI detection is evolving beyond simple metrics to deep learning, crucial for preserving human authorship and combating large-scale deception.
• ‘Obvious markers of AI’: doubts raised over winner of short story prize — The definitive detection of AI-generated creative work remains elusive, posing challenges for literary awards and publishers.
• Beating the Style Detector: Three Hours of Agentic Research on the AI-Text Arms Race — Agentic research rapidly reproduces and extends NLP studies, demonstrating LLMs' ability to evade AI-text detection.
• Tool promises to make lazy academics' AI-written papers sound more human — Academic Humanizer uses AI to refine AI-generated academic drafts, aiming for human-like clarity and voice.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.