AI spots fake reviews, then tracks how they spread

AI detects – A new study in the International Journal of Information and Communication Technology describes an AI-powered system that detects fake online reviews by combining text analysis, reviewer behavior signals, and image recognition. It also uses a Transformer model

The moment you trust a glowing review and get something that’s clearly not what you expected, it sticks. Fake reviews don’t just mislead—they shape what people buy, and they do it using convincing details that current safeguards often can’t catch.

A new study published in the International Journal of Information and Communication Technology sets out to fight back with an AI system designed for two jobs: detecting fake reviews and tracing how they spread.

The researchers point to why many existing fake review detectors stall: they focus heavily on the review text. That worked when fake reviewers relied on copy-paste language. But the study says fake accounts have since gotten smarter. They pair carefully written text with misleading images to make reviews look authentic. Text-only tools. the researchers argue. struggle when deception shows up in the pictures as well—and that leaves both shoppers and honest sellers exposed.

To address that, the system they built doesn’t gamble on a single clue. It analyzes the review text in two ways at the same time. One component is a text convolutional neural network, designed to capture surface-level patterns in the words. The other uses pre-trained language models to capture deeper meaning. The system also looks beyond the writing to reviewer behavior. because fake accounts often come with default profile pictures and system-generated usernames. while real users are more likely to personalize their accounts.

But the breakthrough here is that it doesn’t stop at text. Review images are analyzed separately using a residual network, a deep learning model commonly used for processing visuals. After the text signals and the image signals are gathered. the system fuses them together to decide whether a review is genuine.

Detection is only the first step. When the system flags a review as fake, a Transformer model kicks in to map its origin and track how far it spread through the network.

On a large dataset from JD.com. the results are reported as strong: the system achieved a recognition accuracy of 94.2% and a tracing accuracy of 93.5% across the methods it was compared against. The researchers say accuracy like this could eventually lead to fewer misleading reviews and more trustworthy ratings for shoppers to rely on.

fake reviews detection AI-powered system JD.com dataset text convolutional neural network pre-trained language models residual network Transformer model review fraud online shopping safety