Skip to main content

AI detectors fooled by style mimicry, study finds

AI text detectors are pretty good at spotting run-of-the-mill machine-written content—but throw in a little style mimicry, and they start to stumble. That's the takeaway from a new study by Epoch AI, which tested three popular detectors and found that nearly 30% of AI-generated texts slipped through when the models were asked to copy a specific author's voice.

The research team put Pangram (version 3.3.2), GPTZero (2026-05-11-base), and Originality.ai (Turbo3.0.2) through their paces. They built a test set of 495 original human-written pieces spanning blogs, novels, and scientific writing—all created before ChatGPT launched in November 2022 to avoid any training data contamination.

Near-perfect on standard AI text

When the detectors faced ordinary AI-generated content, they performed admirably. The average false negative rate—where AI text is mistakenly labeled as human—was just 0.7%. Pangram and GPTZero also had zero false positives on human text, while Originality.ai flagged 19 human-written pieces as AI-generated, a 3.8% false positive rate.

Image

Style mimicry changes the game

Then things got interesting. The researchers had three large language models—Claude Opus4.8, GPT-5.5, and Gemini3.1Pro—study five real works by an author and generate new content in that same style. Out of 297 style-mimicking texts, about 13% on average escaped detection. Pangram missed 10%, GPTZero 11%, and Originality.ai 18%.

Scientific writing: the weak spot

The biggest blind spot? Academic writing. For scientific-style AI text, false negative rates jumped to 25% for Pangram, 24% for GPTZero, and 29% for Originality.ai. Some model combinations fared even worse: Pangram missed 48% of Gemini-generated academic texts, and Originality.ai missed 39% of GPT-5.5's academic output.

Despite using different technical approaches—neural networks, text predictability analysis, statistical pattern recognition—all three detectors showed similar vulnerabilities. The study suggests that as large language models get better at mimicking human writing, current detection technology still has a long way to go, especially in fields like education and research where authenticity matters most.

Key Points

  • Standard AI text detection is highly accurate: False negative rate of just 0.7% for ordinary AI content.
  • Style mimicry reduces accuracy: About 13% of imitative texts evade detection, with some detectors missing up to 18%.
  • Scientific writing is hardest to detect: False negative rates reach 29% for academic-style AI text, and up to 48% for certain model-detector combinations.
  • All detectors share similar weaknesses: Despite different technologies, they all struggle with style imitation, especially in scientific contexts.