Latest research from Epoch AI shows that mainstream AI text detectors can almost perfectly identify ordinary AI-generated content, but when large language models deliberately imitate the writing style of specific authors, detection accuracy significantly decreases, with scientific writing being the most difficult scenario 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: 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 false negative rate of the three detectors was as high as only 0.7%. In human text identification, Pangram and GPTZero did not have any false positives, while Originality.ai falsely identified 19 human texts as AI-generated, with a false positive rate of 3.8%.
The study further had Claude Opus4.8, GPT-5.5, and Gemini3.1Pro learn five real works provided by the author and generate new content in the same writing style. Among 297 style-mimicking texts, on average about 13% were not detected by the detectors, with Pangram's false negative rate at 10%, GPTZero at 11%, and Originality.ai at 18%.
Scientific writing became the weakest area for detector performance. The false negative rates of Pangram, GPTZero, and Originality.ai for academic-style AI text reached 25%, 24%, and 29%, respectively. In some model combinations, the performance was even worse, such as Pangram missing 48% of academic texts generated by Gemini, and Originality.ai missing 39% of academic content generated by GPT-5.5.
Despite using different technical approaches such as neural networks, text predictability analysis, and statistical pattern recognition, all three detectors exposed similar shortcomings. The study shows that as large models become increasingly skilled at imitating human writing styles, existing AI text detection technologies still face significant challenges in scenarios that rely on the authenticity of academic writing, such as education and research.
