Detection of AI-Generated and Deceptive Fake Reviews on E-Commerce Platforms Using Lightweight Machine Learning
Product reviews shape what people buy, yet their credibility is under a fresh threat: large language models (LLMs) can now mass-produce reviews that are fluent, varied and grammatically flawless. Current detectors cope with this only in part. The best-performing published systems lean on transformer models, paid LLM interfaces or reviewer metadata, resources that smaller platforms often lack, whereas lighter classical models were largely designed around older, human-written spam. We compare five recent studies of fake-review detection, looking at their data, labelling, methods, reported accuracy and acknowledged limitations. Two patterns emerge: accuracy tends to rise together with computational cost, and some of the highest accuracies rest on heuristic or weakly supervised labels, which casts doubt on what they actually measure. Drawing on these observations, we propose a lightweight, interpretable pipeline that pairs TF-IDF n gram features with stylometric indicators of machine generated text and compares four classifiers: Naïve Bayes, Logistic Regression, Support Vector Machine and Random Forest. The accompanying evaluation protocol is fully specified and includes stratified hold-out and cross validation, cross-generator and cross-domain tests, paraphrase robustness and ablation. Experimental results for the proposed pipeline are not included in this version, and no performance figures are claimed for it. The system is intended as a risk-scoring aid for human moderators, never as proof of AI authorship.
Authors
- Snehal Patil
- Arpita Singh (ORCID: https://orcid.org/0000-0001-8561-587X)
- Mahima Wankhade
- Sanyogita Dahale
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24370422
- Primary Topic
- Spam and Phishing Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00