Epoch AI tested three major AI text detectors: Up to 30% of the content was missed after imitating human writing style. Epoch AI’s latest research shows that mainstream AI text detectors can almost perfectly identify ordinary AI-generated content. However, when large language models deliberately imitate the writing style of a specific author, the detection accuracy drops significantly, and scientific writing becomes the most difficult to identify.

Epoch AI tested three major AI text detectors: Up to 30% of the content was missed after imitating human writing style. Epoch AI’s latest research shows that mainstream AI text detectors can almost perfectly identify ordinary AI-generated content. However, when large language models deliberately imitate the writing style of a specific author, the detection accuracy drops significantly, making scientific writing the most difficult scene to identify. The research team tested three mainstream AI text detectors: Pangram (3.3.2), GPTZero (2026-05-11-base) and Originality.ai (Turbo3.0.2), and built a test set containing 495 original human texts, covering three categories of blogs, novels and scientific writing. All samples were created before ChatGPT was released in November 2022 to avoid training data contamination. The test results showed that for ordinary AI-generated text, the highest missed detection rate of the three detectors was only 0.7%; in terms of human text recognition, Pangram and GPTZero had no false positives, while Originality.ai mistakenly determined that 19 human texts were generated by AI, with a false positive rate of 3.8%. The research further allowed Claude Opus4.8, GPT-5.5 and Gemini3.1Pro to learn 5 real works provided by the authors and generate new content according to the same writing style. Among the 297 style imitation texts, an average of about 13% were not recognized by the detector, of which Pangram’s missed detection rate was 10%, GPTZero’s was 11%, and Originality.ai’s was 18%. Scientific writing emerged as the area where the detector performed weakest. The missed detection rates of Pangram, GPTZero and Originality.ai for academic-style AI texts reached 25%, 24% and 29% respectively. The performance of individual model combinations is even worse. For example, Pangram missed 48% of the academic text generated by Gemini, and Originality.ai missed 39% of the academic content generated by GPT-5.5. Although the three detectors use different technical routes such as neural networks, text predictability analysis and statistical pattern recognition, they all expose similar shortcomings. Research shows that as large models become more and more adept at simulating human writing styles, existing AI text detection technology still faces greater recognition challenges in scenarios that rely on the authenticity of academic writing, such as education and scientific research. via AI News (author: AI Base)