The emergence of advanced generative AI models, such as GPT, Gemini, and Grok, has increasingly blurred the distinction between AI-generated and human-written text. In the context of Vietnamese news articles, distinguishing these sources is crucial to combat misinformation and uphold journalistic integrity. This study proposes a novel model architecture that integrates Multilingual-E5 embeddings with Transformer-based models, including BERT-base, RoBERTa-base, DistilBERT, and DeBERTaV3-base, to effectively classify text as either AI-generated or human-authored. We curated a balanced dataset comprising 200,000 Vietnamese news articles: 100,000 sourced from reputable outlets such as Thanh Niên and VnExpress, and 100,000 generated by advanced large language models (LLMs) like GPT-4o Mini and Gemini Flash 1.5, ensuring diversity in both content and style. Using the robust semantic representations of Multilingual-E5 alongside the powerful feature extraction capabilities of the Transformer models, our architecture achieves a superior classification accuracy that exceeds 99 % across various configurations. This shows its ability to detect subtle textual nuances effectively. These findings confirm the feasibility of building high-precision AI-human text classification systems customized for Vietnamese, offering scalable solutions to the field. This work establishes a strong foundation for AI text detection and holds promise for adaptation to other languages.