AI Text Detection with RoBERTa-BiGRU and Double Multi-Head Attention

The rapid proliferation of AI-generated content has raised pressing concerns in domains such as education, publishing, and digital media. To address the challenges of detecting semantically invariant and humanized AI text, this paper introduces RoBERTa-BDMA, a novel architecture that integrates RoBERTa-based contextual embeddings, BiGRU-driven bidirectional sequence modeling, and a dual multi-head attention mechanism for hierarchical feature extraction. Extensive experiments on the HC3 and HC3-Plus datasets demonstrate its effectiveness, achieving 98.76% accuracy on the question-answer dataset and 92.42% on the semantic-invariant dataset. It also maintains $\mathbf{9 1. 3 \%}$ detection confidence on paraphrased inputs and 89.5% accuracy under natural noise conditions, including typographical errors and code-switching. Compared with DetectGPT and GPTZero, RoBERTa-BDMA offers enhanced semantic discrimination and greater adaptability to structurally varied or humanized inputs, making it a more effective solution in supervised detection scenarios. Future research will focus on expanding evaluation to multilingual, noise-rich corpora and exploring lightweight model optimization techniques.