This article shares a custom prompt designed specifically to reduce the detection rate of AI-generated papers. In actual tests, this prompt successfully lowered the AI rate to 6% in Weipu detection and to nearly 0% in Tencent Zhuque detection. The author conducted tests using the Gemini model, where the 2.5 Pro version produced higher quality output, while the 2.5 Flash version showed quality deviations. It is recommended that users generate content in small sections (no more than 1000 words at a time) and clear the context to maintain a low detection rate. Additionally, the author points out that the paperyy tool intentionally inflates the AI rate to induce paid paraphrasing services, whereas the speedai research assistant shows a detection rate of about 14%, which can be used as a reliable reference. This experience provides practical paraphrasing techniques for academic writing, helping researchers cope with the challenges of AI content detection.
