AI Is Now Writing Fake Reviews — And Catching Them Too

A team of researchers from Guizhou University of Finance and Economics, Harbin Institute of Technology (Shenzhen), and Guizhou University of Commerce combed through 211 studies published from 2018 to early 2026 — all about how to spot fake reviews online. Their paper, accepted by the journal Information Fusion, reads like a field guide to a problem that quietly affects what you buy, where you eat, and which hotel you book.

The problem got much harder in the last few years. Older fake reviews were often clumsy — odd phrasing, repeated templates, suspicious timing. Today's AI language models can churn out fluent, context-aware reviews that sound like your neighbor wrote them. The good news: the same technology that writes convincing fakes can also read them. Newer detection systems understand meaning, not just keywords, so they can flag language that feels slightly off even when the grammar is perfect.

The paper's biggest practical insight is that no single clue is enough. The strongest detectors combine several kinds of evidence: the review text itself, the emotion behind it, the star rating given, when it was posted, and who reviewed what before. Some systems also look at photos and videos, or pull in outside knowledge about a business. Most studies test themselves on well-known datasets from Amazon, Yelp, and a review site called OpSpam.

But there's a measurement problem. Different studies build their training data differently and split it differently, so their accuracy numbers can't be fairly compared. Many open challenges remain: fake-review writers adapt to detectors, a system trained on restaurant reviews may flop on hotels, and missing data can break things entirely. The authors also call for clearer explanations from these systems and a more trustworthy way to judge them.