Distinguishing between human-written and AI-generated text has become increasingly challenging with the rapid advancement of large language models. This paper presents an ELECTRA-based deep learning framework for binary authorship verification, leveraging replaced-token discrimination to capture token-level irregularities characteristic of AI-generated content. The proposed system fine-tunes the google/electra-small-discriminator model using GPU-optimized training with mixed-precision and gradient accumulation for computational efficiency. Experiments were conducted on a balanced dataset comprising human-authored and AI-generated texts, with evaluation performed using accuracy, precision, recall, and F1-score. The framework achieves strong performance and demonstrates robustness across unseen generative models, indicating effective generalization beyond training distributions. These results highlight the suitability of ELECTRA-based discriminative modeling for scalable and reliable authorship verification in modern AI-driven text generation environments.