Nankai team develops breakthrough system to detect AI-generated text

A research team at Nankai University in Tianjin has made major strides in detecting text generated by artificial intelligence, developing a system that significantly reduces false positives and false negatives, a problem that has plagued many existing tools.

The team's research paper on the system has been accepted by ACM Multimedia 2025, one of the world's leading computer science conferences. Their detection feature is now integrated into Paper-Mate, an AI research assistant developed by Nankai professors Li Zhongyi and Guo Chunle, and is available free of charge.

The system has more than 1,000 monthly active users, including teachers and students from several universities such as Peking University, Zhejiang University and Sun Yat-sen University, said team member Fu Jiachen.

"Many users have provided feedback, indicating that PaperMate outperforms similar tools on the market in terms of false positives and false negatives, offering more accurate and reliable detection results," Fu said.

"Current AI detection tools for academic papers often falsely accuse authors. For instance, my senior, while writing his thesis, used existing AI detection tools and found that some of his original content was mistakenly flagged as AI-generated," he said.

Explaining the reasons behind misidentifications by current AI text detection methods, Fu said: "If we liken AI text detection to an exam, the training data of the detector is akin to daily practice questions. Existing detection methods tend to mechanically memorize fixed routines for answering questions, and their accuracy drops significantly when faced with entirely new problems.

"In theory, to achieve universal detection, we would need to train on data from all major models, which is nearly impossible given the rapid iteration of these models today."

Enhancing the detection's generalization ability and enabling the detector to apply principles across various scenarios is crucial for improving AI text detection performance.

The Media Computing Laboratory of Nankai University's School of Computer Science has not only revealed the performance limitations of existing AI detection methods from an evaluation perspective, but has also proposed a "Direct Discrepancy Learning" optimization strategy.

This strategy teaches AI to discern between human- and machine-generated text, achieving a significant breakthrough in detection performance.

"In essence, we improve the accuracy of the detection algorithm to reduce the false positive rate," Fu said.

He also introduced MIRAGE, the team's benchmark dataset.

"We collected human-written texts and then had AI large models refine these texts, resulting in a set of human-original texts and AI-generated texts. By applying both existing algorithms and our algorithm to these texts, MIRAGE records the detection accuracy," he explained.

The dataset test results show that the accuracy of existing detectors dropped from 90 percent to around 60 percent, while detectors trained with Direct Discrepancy Learning maintained an accuracy of more than 85 percent. Compared to Stanford University's DetectGPT, performance improved by more than 70 percent, he said.

Li Zhongyi, who heads the Media Computing Laboratory at Nankai University's School of Computer Science, said the findings highlight fundamental flaws in many current detection systems and offer a practical path forward.

"With AI-generated content developing so rapidly, we will keep iterating our technology and benchmark to achieve faster, more accurate and cost-effective detection," Li said. "Our goal is to use AI itself to make every piece of work shine."

Liu Tingyu contributed to this story.

yandongjie@chinadaily.com.cn