A new academic survey takes stock of eight years of research into catching fake online reviews, right as the tools for faking them get much better.
The paper reviews 211 studies published between 2018 and early 2026, tracing fake review detection from early machine learning and deep learning methods through today's pre-trained language models and large language models. It organizes the work by what evidence detectors actually use: review text, sentiment, rating patterns, timing metadata, user-product graphs, images and other multimodal content, external knowledge, and signals specific to LLM-generated text. The authors also compare reported results across the standard Amazon, Yelp, and OpSpam benchmarks, while flagging that inconsistent labeling and evaluation setups make those comparisons shakier than they look.
The real story here is the arms race baked into the abstract: the same LLMs that write more convincing fake reviews are also the best tool we have for catching them. That is not a stable equilibrium. It means detection systems are chasing a moving, self-improving target rather than a fixed set of scam tactics.
The survey's list of open problems is the more honest part of the paper. Adversarial generation, cross-domain transfer, and trustworthy evaluation of AI-written deception are all unsolved. Translation: nobody has a reliable answer yet, and platforms betting on automated detection alone should not get comfortable.