Detection of AI-Generated Fake Reviews: A Comparative Analysis of Transformer and Recurrent Architectures Across Language Model Generations

Bade, Sai Kalyan (2025) Detection of AI-Generated Fake Reviews: A Comparative Analysis of Transformer and Recurrent Architectures Across Language Model Generations. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (813kB) | Preview

Abstract

The increased utilization of advanced language models has enabled the creation of various fake product reviews, which could adversely affect consumer trust and the credibility of e-commerce platforms. The proposed work investigates the detection of AI-written fake reviews by comprehensively evaluating four detection models on various generations of language models and product types. The study evaluated the Attention-GRU, RoBERTa, DeBERTa, and XLNet models on two datasets: the reviews written by the GPT-2 across ten Amazon products, including FraudSquad reviews written by the three modern language models Qwen-72B, LLaMA-3-8B, and DeepSeek-R1-Qwen across seven categories. The experimental result revealed that the current fine-tuned language models produce more detectable content than their predecessors, GPT-2, which negates the presumptive hypotheses stating the greater difficulty level of detection represented by newer LLMs. Transformer models have always given more accurate results than the recurrent models. For instance, DeBERTa and RoBERTa achieved 99-100% accuracy against modern LLM content, while Attention-GRU achieved only 88-96%. These models of detection suggested good generalization when there was a significant difference in the categories with the total error of only 2 to 5 points between the two kinds of products. The paper is the first to systematically compare the ability of different generations of language models to detect fake reviews gained through AI and corroborates the capacity of current transformer-based detection models to detect high-quality fake reviews with near-perfect precision, thus offering valuable lessons to approaches with e-commerce platform protection.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Moldovan, Arghir Nicolae UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing H Social Sciences > HF Commerce > Electronic Commerce Z Bibliography. Library Science. Information Resources > ZA Information resources > ZA4150 Computer Network Resources > The Internet > World Wide Web > Websites > Online social networks T Technology > TK Electrical engineering. Electronics. Nuclear engineering > Telecommunications > The Internet > World Wide Web > Websites > Online social networks
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	07 Sep 2026 09:08
Last Modified:	07 Sep 2026 09:08
URI:	https://norma.ncirl.ie/id/eprint/9848

Actions (login required)

View Item