Work

Detecting AI-Generated Text Using a Hybrid Deep Learning Framework for Authenticity Verification

Year: 2025
Type: dissertation
Abstract: The increasing prevalence of AI-generated text across digital platforms has created significant challenges for maintaining authenticity, preventing misinformation, and ensuring trustworthy communication. This study addresses the need for reliable detection of human-written and LLM-generated content by developing a lightweight hybrid architecture that integrates ALBERT’s parameter-efficient contextual embeddings with the sequential learning capability of an LSTM network. Using a curated Human vs LLM dataset comprising text from Human, GPT-3.5, Claude-v1, and LLaMA-30B sources, the methodology follows the CRISP-DM framework, incorporating systematic data preparation, text cleaning, exploratory analysis, and model construction. ALBERT embeddings capture deep semantic and syntactic relationships within text, while the LSTM layer learns temporal and stylistic patterns characteristic of human and model-generated language. Evaluation using macro-average metrics demonstrates that the hybrid ALBERT+LSTM architecture achieves superior performance, obtaining an accuracy of 0.91 along with balanced precision, recall, and F1-score values of 0.91 across all classes. These results highlight the model’s strong generalization capability and its ability to capture subtle linguistic distinctions even in short, informal social media posts. The study demonstrates the suitability of transformer–recurrent hybrid architectures for real-world text authenticity detection and emphasizes their potential for scalable deployment in environments where rapid and accurate identification of AI-generated content is essential.
Language: English
Open Access status: green
OpenAlex ID: W7212334168
Cites: 15
Cited by:
Related to: 10