Dual-Supervision Collaborative Machine-Generated Text Detection Method

With the rapid development of Artificial Intelligence Generated Content (AIGC) technology, its impact on the authenticity and reliability of information has become increasingly significant, leading to a series of technical and ethical issues such as data pollution, unclear ownership, and social trust crises. Detecting machine-generated texts can effectively curb the spread of false information and improve data quality. However, humanmachine hybrid texts (such as machine-completed or machine-polished texts) are semantically similar to human-written texts, making existing detection methods prone to misclassification. To address this issue, a dual-supervised collaborative machine-generated text detection method is proposed. This method employs a multi-class contrastive loss to increase inter-class feature distances and uses edit distance to measure the machine involvement degree in text generation. Combined with class labels, both are used as supervisory signals to train the model, enhancing feature differentiation among different text types and improving detection accuracy. Experimental results on multiple real datasets demonstrate that the method achieves optimal detection accuracy in complex text scenarios.