Fake Review Detection Survey: The Arms Race from Pre-trained Models to Large Language Models

Online reviews have long become a critical information channel shaping consumer decisions, platform governance, and corporate reputation. When a single review can influence purchasing behavior, recommendation rankings, or public trust, the damage that fake reviews inflict on the broader information ecosystem becomes all the more severe. A recent survey covering 211 research papers systematically maps the evolution of fake review detection — from traditional machine learning to the era of large language models (LLMs) — and reorganizes the field's research landscape through the distinctive lens of "information fusion."

LLMs Turn the Problem into a Double-Edged Sword

The rise of large language models has pushed the long-standing problem of fake review detection in two diametrically opposite directions. On one hand, LLMs can generate fluent, contextually convincing deceptive reviews that are harder to spot than ever — no longer riddled with grammatical errors or obvious template traces, they can mimic real users' tone, specific details, and even emotional nuance. On the other hand, pre-trained language models (PLMs) and LLMs also provide the detection side with stronger semantic representation capabilities, enabling models to capture deeper textual anomaly signals.

This "shared origin" between attack and defense is the most fundamental tension in the field today. Both the generation side and the detection side draw on the same class of underlying technology, meaning detection systems must continuously contend with adversarial content driven by increasingly capable models. The survey notes that this dynamic arms race has gradually rendered traditional surface-feature-based methods obsolete, driving the research focus toward semantic understanding and multi-source evidence fusion.

The advantages of PLMs and LLMs on the detection side are rooted in the deep semantic representations they acquire from training on massive text corpora. BERT-style PLMs learn to capture fine-grained contextual word meanings through masked language modeling; autoregressive LLMs like the GPT family go further in modeling long-range linguistic patterns. This allows detection models to move beyond surface features like "typo frequency" or "exclamation mark count" and instead identify subtler anomalies such as "emotional tone inconsistent with star rating" or "product descriptions mismatched with category." Yet the generation side benefits equally from these advances — reviews generated by GPT-4, Claude, and similar models now closely resemble genuine user reviews in fluency, detail richness, and emotional naturalness, leaving style-based detectors with almost nothing to work with. This pattern — where attack and defense share the same technical foundation — is referred to in adversarial security research as an "arms race," and its core paradox is that the detection side cannot assume it has access to more advanced models than the generation side.

Remapping the Research Landscape by "Evidence Source" and "Fusion Level"

One of the most valuable contributions of this survey is its proposal of an organizing framework centered on information fusion. Rather than simply listing papers chronologically or by method, it structures existing work according to evidence sources and fusion levels.

The evidence sources identified by the survey include: review text itself, sentiment signals, rating behavior, temporal metadata, user–product relationship graphs, multimodal content, external knowledge, and LLM-generated signals. This classification reveals an important insight — signals from any single dimension are often insufficient to reliably identify fake reviews. Text can be disguised, but anomalous rating behavior, concentrated posting times, or suspicious user network structures may expose the deception.

The survey traces the full arc of technical development: from traditional machine learning and deep learning, to PLM-based methods, to LLM-based methods, with a focus on how different approaches combine textual, behavioral, structural, and multimodal evidence. This fusion perspective provides a clear explanatory framework for understanding why certain methods are more robust.

Information fusion is a concept originating in sensor engineering that refers to integrating signals from multiple heterogeneous data sources into a judgment more reliable than any single source alone. In the context of fake review detection, fusion levels are typically organized into three tiers: feature-level fusion (concatenating raw features from different sources and feeding them into a unified classifier), score-level fusion (each sub-model scores independently and the scores are weighted and combined), and decision-level fusion (multiple classifiers make independent judgments, with a final conclusion reached via voting or meta-learning). Each level involves trade-offs in flexibility, interpretability, and robustness to missing data. For example, feature-level fusion allows the model to learn cross-source interaction signals, but if any one source is missing, the entire pipeline breaks down; decision-level fusion naturally supports "inference from a subset of available sources," directly relating to the open challenge of "missing-source robustness" discussed in the survey.

The Progress and Pitfalls of Benchmarking

On the performance side, the survey analyzes reported result trends across widely used benchmark dataset families including Amazon, Yelp, and OpSpam. Overall, detection performance has genuinely improved as model capabilities have advanced.

But the authors also sound an important warning: the numbers across different studies are not necessarily comparable. Differences in label construction pipelines, data splitting strategies, and evaluation protocols mean that apparent performance gains may partially reflect looser evaluation setups rather than genuine methodological progress. This is a sobering message for the entire field — in the absence of unified evaluation standards, cross-paper comparisons of accuracy or F1 scores can be misleading. For practitioners looking to deploy real-world systems, this means not blindly trusting leaderboard numbers and instead focusing on reproducibility and evaluation fairness.

OpSpam, Amazon, and Yelp are the three most widely used benchmark datasets in this field, but their labels come from different sources, meaning the definition of "fake review" they measure is not the same. OpSpam was constructed by Myle Ott et al. in 2011 by crowdsourcing deceptive hotel reviews through Amazon Mechanical Turk — it represents "artificially written, known-fake content." The Yelp dataset relies on the platform's own filter labels, and filtered reviews are not necessarily human-verified. Amazon dataset series typically use extreme ratings from "unverified purchases" as weak supervision labels. These differences mean a model that scores highly on one dataset may perform much worse on another — not because of poor generalization, but because the operational definition of "fake review" is fundamentally different. Understanding this is a prerequisite for correctly interpreting cross-paper numerical comparisons.

Six Open Challenges That Remain Unsolved

The survey concludes by identifying critical open problems that the field has yet to crack — these directions also represent the most promising areas for future research:

• Adversarial generation: How to handle LLM-driven, continuously evolving deceptive content
• Cross-domain transfer: How models trained on one platform or product category can generalize to other settings
• Uncertainty-aware fusion: How to appropriately represent and handle confidence levels when fusing multi-source evidence
• Missing-source robustness: Whether systems can function reliably when certain evidence sources are unavailable
• Interpretability: Whether detection decisions can provide human-understandable justifications
• Trustworthy evaluation: How to build credible evaluation frameworks for AI-generated deceptive content

These six challenges collectively point to a deeper underlying difficulty: in an era of widespread generative AI, how do we build fake review detection systems that are simultaneously accurate, trustworthy, robust, and interpretable?

The cross-domain transfer challenge is especially acute in fake review detection, because user behavioral norms, review styles, and fraud motivations vary systematically across platforms: fake reviews on dining platforms are often mass-posted by competitors hiring "review farms," while fake reviews on e-commerce platforms more commonly involve genuine buyers cooperating under "sales-boosting" incentive schemes. Domain adaptation techniques — including adversarial training, meta-learning, and prompt engineering — are considered potential remedies, but no universally applicable solution has yet been rigorously validated in real cross-platform settings. The interpretability challenge is closely tied to regulatory compliance: regulations such as the EU's Digital Services Act (DSA) require platforms to provide auditable justifications for content moderation decisions, which means black-box detection models may face legal challenges — pushing the research community to direct attention toward transparency methods based on attention visualization, counterfactual explanations, or rule extraction.

Implications for Industry

For e-commerce platforms, content communities, and regulators, this survey sends a clear signal — the era of relying solely on text classification models is passing. Effective fake review governance requires integrating behavioral, structural, temporal, and other multi-dimensional evidence, and system design must grapple with the reality that adversaries have equal access to advanced generative capabilities.

At the same time, as the boundary between AI-generated and human-generated content grows increasingly blurred, the very goal of detection is evolving: from "identifying human-fabricated fakes" to "identifying AI-fabricated fakes." This is not merely a technical problem — it concerns the long-term maintenance of platform governance and public trust. This survey, spanning eight years and over two hundred papers, provides a systematic map for understanding this continuously evolving arms race.