PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Fake Review Detection: From Pre-trained Language Models to Large Language Models

Fanji Yang, Huiyao Chen, Xi Yu, Meishan Zhang, Xiaohong Xiao, Mingsen Deng

arXiv:2609.30292v1cs.CLcs.AI

TL;DR

LLMs make fabricated reviews more fluent and coherent, complicating detection based on conventional signals. This survey organizes detection methods by evidence and fusion, finding that effective approaches combine heterogeneous signals and that the field increasingly centers on semantic deception and evidence reliability.

  • Problem

    LLMs produce fluent, coherent fabricated reviews that can make conventional detection signals unreliable.

    Source: §5.6 · p. 27

  • Method

    The survey organizes fake review detection by evidence source and fusion level, reviews method evolution, and analyzes benchmark trends and evaluation protocols.

    Source: §1 · p. 3

  • Results

    The reviewed trajectory shifts from surface linguistic anomalies toward semantic deception and evidence reliability, while effective approaches combine heterogeneous signals.

    Source: §5.6 · p. 27, §8 · p. 33

  • Takeaways & Limitations

    Future detectors need uncertainty-aware fusion that estimates source reliability and degrades gracefully when evidence sources are missing.

    Source: §7.2 · p. 33

    PLM-based methods can fail on LLM-generated reviews that are semantically coherent and stylistically indistinguishable from authentic writing.

    Source: §4.5 · p. 21

Abstract

from arXiv · show

Online reviews shape consumer decisions, platform governance, and corporate reputation.Fake reviews compromise this information channel by injecting deceptive evidence into rating systems, recommendation pipelines, and public trust mechanisms.The rise of large language models, or LLMs, has changed the problem in two directions.LLMs can generate fluent and context-aware deceptive reviews, while pre-trained language models, or PLMs, and LLMs also provide stronger semantic representations for detection.This survey reviews fake review detection from an information fusion perspective, covering 211 studies published from 2018 to early 2026.We organize existing work by evidence source and fusion level, covering review text, sentiment, rating behavior, temporal metadata, user-product graphs, multimodal content, external knowledge, and LLM-generated signals.We trace the development from traditional machine learning and deep learning to PLM-based and LLM-based methods, and examine how different approaches combine textual, behavioral, structural, and multimodal evidence.We also analyze reported performance trends on widely used Amazon, Yelp, and OpSpam benchmark families, while noting the limitations caused by different label construction procedures, data splits, and evaluation protocols.Finally, we identify open problems in adversarial generation, cross-domain transfer, uncertainty-aware fusion, missing-source robustness, interpretability, and trustworthy evaluation for AI-generated deceptive content.

1. Introduction

Fake reviews threaten consumer trust and platform governance, while LLMs intensify the challenge by generating fluent deceptive content and supporting stronger semantic detection. This survey traces model evolution and organizes research around evidence fusion, benchmark trends, and LLM-era challenges.

1. Introduction

  • More than 70% of customers consult online reviews before purchases, and 63% say reviews directly influence buying behavior; fake reviews also carry estimated annual losses exceeding $152 billion.The introduction frames moderation as a large-scale governance problem requiring automated screening, manual review, and platform accountability.Source: §1 · p. 2
  • Profit-driven merchants, manipulation services, and incentive-motivated users produce fake reviews that spread through recommendation algorithms and social diffusion, amplifying governance challenges beyond text classification.Figure 1 frames sources, generation and dissemination mechanisms, and impacts as parts of the same process.Source: §1 · p. 2, Figure 1 · p. 2
  • Detection evolved from handcrafted linguistic, sentiment, behavioral, and network features to deep representation learning and PLM fine-tuning, while LLMs now support semantic detection and fluent deceptive generation.The survey covers research from 2018 to early 2026 through the joint lens of model evolution and information fusion.Source: §1 · pp. 2–3
  • The survey organizes methods by evidence source and fusion level, covering text, sentiment, behavior, temporal metadata, graphs, multimodal content, external knowledge, and LLM-generated signals.It also reviews how traditional machine learning, deep learning, PLM-based, and LLM-based methods combine textual, behavioral, structural, and contextual evidence.Source: §1 · p. 3
  • The survey compares reported trends on representative benchmarks and identifies open problems in adversarial robustness, cross-domain adaptation, uncertainty-aware fusion, and interpretable decision-making.Its benchmark analysis discusses how label construction, data splits, class imbalance, and evaluation metrics affect comparability.Source: §1 · p. 3

2. Related Work

Prior surveys track the shift from handcrafted features and conventional learning toward deep, graph-based, and foundation-model approaches. The survey positions its contribution as a fusion-oriented account of LLM-era methods and their challenges.

2. Related Work

  • Early surveys reviewed handcrafted linguistic, behavioral, temporal, metadata, and graph features alongside conventional learning methods; later surveys expanded toward deep learning, graph models, and foundation models.Table 1 is presented as a comparison of representative surveys’ technical coverage and research scope.Source: §2 · pp. 3, 6
  • Existing surveys provide useful coverage of methods, datasets, and evaluation protocols, but do not yet offer a comprehensive fusion-oriented account of LLM-era detection in practical applications.Some recent surveys discuss Transformer models and preliminary LLM applications.Source: §2 · p. 6
  • Prior reviews rarely connect model evolution with fusion of textual, behavioral, structural, and contextual signals, and emerging LLM approaches remain fragmented across research directions.The approaches include direct LLM discrimination, semantic feature enhancement, generative discrimination, and LLM-driven augmentation and adversarial learning.Source: §2 · p. 6
  • This survey links evidence sources, fusion levels, and LLM-era detection paradigms, while discussing limitations and research directions for reliable detection under AI-generated deceptive content.It frames machine-generated content as raising issues of source reliability, robustness, generalization, trustworthy evaluation, and adversarial adaptation.Source: §2 · p. 6

3. Background and Problem Formulation

This section defines the survey’s literature corpus and frames fake-review detection as a shift from surface cues toward semantic deception and heterogeneous evidence fusion. It also surveys conventional methods and their limits, dataset construction and labeling, and evaluation metrics for assessing performance, reliability, interpretability, and robustness.

3.1. Survey Methodology

  • Studies are coded by dominant modeling family and evidence source, with fusion classified as feature-, representation-, graph-, decision-level, or uncertainty-aware.The coding supports comparison across modeling paradigms while separately recording each method’s evidence sources.Source: §3.1 · p. 7
  • From 359 records retrieved (2018–March 27, 2026), the representative coded corpus retained 211 studies: 92 ML, 71 DL, 30 PLM, and 18 LLM.The survey describes its collection as representative rather than a strict systematic review; the 2026 count covers only publications available by early 2026.Source: §3.1 · pp. 6–7

3.2. Task Definition and Scope

  • Fake reviews are deliberate deceptive content, classified as untruthful, brand-oriented, or non-reviews by Jindal and Liu, and can closely resemble genuine reviews.This resemblance makes manual inspection and verification difficult.Source: §3.2 · pp. 7–8
  • LLM-generated reviews use fluent, context-aligned narratives, shifting detection from surface anomalies toward semantic deception and manipulative intent [201].Their naturalness can approximate authentic user writing, weakening lexical-statistics and shallow-semantic approaches.Source: §3.2 · pp. 8–9

3.3. Conventional Fake Review Detection

  • Classical methods rely on engineered features and classifiers; unsupervised K-means pseudo-labeling reached 88.1% average accuracy on JD.com electronics, but degraded on reviews under 20 words.Earlier methods also explored feature selection and ensembles, with stacked ensembles outperforming simple voting ensembles but remaining below deep-learning models.Source: §3.3 · pp. 9–10
  • Traditional methods’ manual features and surface focus miss semantic intent and long-term identity-mimicking, while evolving spam strategies cause performance degradation over time on Yelp.Behavioral signals can also be sparse and difficult to interpret when user identities are anonymous or data are obfuscated.Source: §3.2 · p. 8, §3.3 · pp. 9, 10, 12
  • Deep networks learn representations from raw text: CNNs capture local features, recurrent models capture sequence information, and attention can fuse linguistic and behavioral signals.Hybrid CNN-recurrent architectures and hierarchical user- and product-level attention extend these capabilities.Source: §3.3 · p. 12
  • Knowledge graphs, social relations, and reviewer or merchant credibility add structural and behavioral context to textual evidence for detecting complex manipulation.Dynamic knowledge graphs encode evolving review, product, and user relations; graph models capture coordination, while credibility and activity sequences provide additional signals.Source: §3.3 · pp. 12–13
  • Conventional methods combine knowledge graphs, GANs, graph neural networks, and linguistic, behavioral, or structural features to detect organized behavior and address data scarcity or cold starts.Examples include FakeGAN’s semi-supervised generation, GAN-based cold-start methods, GNN detection of spam reviewers, and multi-feature or collective-behavior models.Source: §3.3 · pp. 13–14
  • CNN- and LSTM-based detectors handle extreme sentiment or templated reviews but struggle with semantic deception, coordinated campaigns, and reasoning about user intent, motivating PLMs.Content-only models miss anomalous correlations among review targets, and their syntactic-level semantics lack reasoning capacity and world knowledge.Source: §3.3 · p. 14

3.4. Datasets and Evaluation Metrics

  • Dataset construction uses filtering algorithms, crowdsourcing, manual annotation, and rules, with differing assumptions that affect label reliability, robustness, generalization, and fair comparison.The survey organizes representative datasets by scale, domain, available information, application, construction strategy, dataset bias, and benchmark considerations.Source: §3.4 · p. 14
  • Platform-filtered Yelp datasets provide large-scale labels, but proprietary filtering criteria are undisclosed, limiting scrutiny of how reviews are labeled.Examples include Yelp CHI, Yelp-NYC and Yelp-ZIP, and a Chinese dataset.Source: §3.4 · p. 14
  • Crowdsourced datasets use deliberately generated fake reviews but differ from platform-review distributions, while manual labels are costly and can contain noise.AMT datasets include 800, 1,600, and 3,032 reviews [203]; manual datasets include 6,000 reviews with 1,398 labeled fake and 3,000 with 712 fake.Source: §3.4 · p. 14
  • Rule-based labeling scales efficiently but can misclassify legitimate repeat reviews as deceptive when its duplication assumptions do not hold.Jindal and Liu used similarity above 0.9 to label reviews in a 5.8-million-review dataset, about 55,000 of them deceptive.Source: §3.4 · pp. 14, 17
  • Because genuine reviews often dominate, classification assessment should emphasize Precision, Recall, and F1 alongside Accuracy, AUC-ROC, True Negative Rate, and Matthews Correlation Coefficient.Precision reflects the reliability of fake predictions, Recall the share of fakes identified, and F1 their harmonic balance; AUC-ROC compares separability across thresholds, while MCC incorporates the full confusion matrix.Source: §3.4 · p. 17
  • Evaluation also considers LLM prediction stability and uncertainty, feature-level explanations, and whether fused detectors remain reliable when sources are missing, noisy, or corrupted.Repeated-trial consistency, confidence variation, and Cohen’s Kappa assess reliability; SHAP supports local and global explanations, while ablations, calibration, and missing-source tests probe robustness.Source: §3.4 · p. 17

4. PLM-Based Fake Review Detection Methods

PLM-based detection progresses from contextual semantic representations to task-fine-tuned classifiers and systems that fuse text with other evidence. The survey reports strong benchmark results while noting constraints involving realistic AI-generated reviews, domain transfer, labeled data, interpretability, and deployment costs.

4.1. PLMs and Semantic Modeling Mechanisms

  • Transformer PLMs provide contextual representations for review semantics; variants and knowledge-enhanced models address different modeling needs, while adaptation strategies include multi-task learning, domain-adaptive pre-training, and augmentation.The survey describes global attention for long-range dependencies, dynamic contextual embeddings, BERT’s bidirectional encoding, and ERNIE’s use of structured knowledge [55, 56, 60, 64]. Domain adaptation can support cross-domain transfer but requires large corpora and computational resources [54].Source: §4.1 · pp. 18–19

4.2. PLMs as Semantic Encoders

  • As semantic encoders, BERT representations reached F1 approximately 0.88 on Yelp, versus approximately 0.82 for CNNs and 0.84 for LSTMs; Amazon RoBERTa results exceeded 0.90 in Accuracy and F1.Static PLM features can reduce training costs, but semantic encoding does not fully exploit task adaptation and remains limited for subtle, context-dependent deception.Source: §4.2 · p. 19

4.3. PLMs as Fine-tuned Classifiers

  • On Yelp, fine-tuned RoBERTa exceeded 0.93 in Accuracy and F1, improving approximately 2 to 3 percentage points over fine-tuned BERT [214].Fine-tuned BERT also achieved Accuracy above 0.90 and F1 from approximately 0.88 to 0.91 on OpSpam; Chinese and low-resource fine-tuning reported F1 ranges of 0.88–0.91 and above 0.85, respectively [227] [203].Source: §4.3 · p. 20
  • Fine-tuning updates PLM parameters for review authenticity classification, and reported results include strong performance on OpSpam, Dianping, Yelp, Chinese, and low-resource language data [227] [203] [214].The approach jointly optimizes semantic modeling and discriminative learning rather than relying on frozen representations.Source: §4.3 · pp. 19–20

4.4. PLMs as Feature Fusion Modules

  • Fusion combines PLM text representations with sentiment, reviewer behavior, visual, metadata, or graph evidence; reported studies describe gains over text-only approaches and detection of natural-sounding reviews with anomalous behavior.Examples include a Yelp behavior-fusion F1 of 0.92 and multimodal F1 scores above 0.90; graph-based methods also target coordinated groups and structural signals.Source: §4.4 · p. 20
  • Feature fusion can improve robustness, but increases model complexity and demands more data collection, annotation, and computational resources.Graph-based constraints may also lose effectiveness where entity relationships are sparse.Source: §4.4 · p. 21

4.5. Limitations of PLM-Based Methods

  • PLMs may miss coherent LLM-generated reviews and lose 5 to 15 percentage points of F1 under cross-domain evaluation.The survey attributes these limits to generated text avoiding learned statistical patterns and to differences in domain vocabulary, style, and contextual logic.Source: §4.5 · p. 21
  • Long-input constraints, dependence on labeled data, and black-box decisions limit PLM deployment in lengthy-review, low-resource, governance, and risk-management settings.DeBERTa partly alleviates long-text issues, while semi-supervised and domain-adaptive methods add training complexity; attention visualization and probing do not provide rigorous explanations [39, 40, 41]. These limitations motivate integrating reasoning, external knowledge, and adaptive learning [42].Source: §4.5 · p. 21

5. Large Language Models in Fake Review Detection

This section examines how LLMs both enable more convincing fake reviews and support detection through semantic reasoning, surveying approaches that combine LLM capabilities with structured and heterogeneous evidence. It also notes concerns about interpretability, reliability, cross-domain generalization, and synthetic-data validity.

5.1. Evolution and Core Mechanisms of LLMs

  • LLMs can generate fabricated reviews that closely mimic authentic user writing, making detection harder, while their semantic reasoning can identify subtle linguistic anomalies and infer manipulative intent.Source: §5.1 · p. 22

5.2. LLM-based Direct Discriminative Detection Methods

  • Unsupervised ChatGPT-3.5, GPT-4, and LLaMA2 detection achieved below 45% accuracy, versus over 96% for fine-tuned DeBERTa-v3 on Amazon and hotel reviews.Cross-lingual performance also degrades, especially for complex emotional expression or implicit semantics [162]. Prompt-generated explanations and features can improve classification; LLMIC reported accuracy above 93% and a 95.9% F1 score on the hotel dataset.Source: §5.2 · p. 23

5.3. LLM-based Feature Extraction and Semantic Enhancement Methods

  • LLM embeddings can serve as features or graph node representations fused with behavioral evidence, and graph-based fake-review experiments improved Precision, Recall, and F1 over text-only models [166].Adjacent-task findings are methodological evidence, not fake-review benchmark results; embedding methods also face substantial computational costs, distribution shifts, and reliance on synthetic training data [243].Source: §5.3 · pp. 23–24

5.4. Generative Discriminative Methods from GANs to LLMs

  • Generative-discriminative approaches model review-text distributions through adversarial learning, language-model likelihoods, and staged generation/classification, supporting detection of covert, anomalous, and fluent fake reviews, including those resembling genuine ones.Persistent challenges include training instability, high computational costs, and the absence of widely accepted unified evaluation standards.Source: §5.4 · p. 25

5.5. LLM-based Data Augmentation and Adversarial Learning Methods

  • LLM-augmented training increased detection accuracy by approximately 0.3 to 10.9 percentage points on DeRev, Amazon, Yelp, and DianPing [77].Generated-data quality depends on prompt design and model performance, while noisy augmentation or poorly designed adversarial perturbations can introduce misleading training signals [77].Source: §5.5 · p. 26

5.6. Limitations and Synthesis

  • Across the reviewed approaches, the field is moving from black-box LLM discriminators toward components in multi-source frameworks combining textual, behavioral, structural, and external-knowledge signals.The surveyed categories include prompt-driven discrimination, semantic representation and enhancement, generative discrimination, and data augmentation with adversarial learning.Source: §5.6 · pp. 26–27
  • LLMs generate fluent, authentic-like fake reviews that weaken conventional signals, yet can improve detection through semantic reasoning integrated with graph modeling, contrastive learning, and knowledge-enhanced architectures.The co-evolution of generation and detection creates a challenge for frameworks that reason across multiple evidence sources.Source: §5.6 · p. 27
  • Effective approaches move beyond text alone by combining reviewer behavior, product-user graphs, external knowledge, and multimodal features with LLM semantic modeling and structural reasoning.Source: §5.6 · p. 27
  • LLM-based detection remains limited by poor interpretability, potentially hallucinated chain-of-thought reasoning, degraded cross-domain generalization under distribution shift, and uncertain ecological validity of synthetic training data.These limitations raise reliability concerns for high-stakes deployment.Source: §5.6 · p. 27

6. Comparative Analysis and Practical Insights

The comparison traces reported benchmark patterns across four families, contrasts ML-, DL-, PLM-, and LLM-based approaches, and considers how evidence fusion addresses emerging deception and multimodal content.

6.1. Performance Evolution on Representative Benchmarks

  • Amazon reports generally rise across model families, but year-to-year variation and differing protocols prevent treating them as a controlled time series.Recent PLM- and LLM-based reports reach Accuracy close to or above 0.95 and F1-score approaching 0.98 in some settings.Source: §6.1 · p. 27
  • OpSpam reports remain high across model generations, with ML-based Accuracy reaching 0.97 in 2019 and later PLM- and LLM-based values close to 0.99 in some configurations.The small reported margin between older and newer methods suggests limited benchmark discrimination, subject to protocol differences.Source: §6.1 · p. 27
  • Yelp-CHI/NYC/ZIP reports fluctuate, and high Accuracy can coincide with lower F1-score, making Accuracy alone incomplete when deceptive reviews are a minority class.Recent PLM F1-score gains are less consistent than those reported on Amazon or OpSpam; domain variation motivates careful cross-domain and behavioral evaluation.Source: §6.1 · p. 28
  • Yelp-Hotel/Restaurant/Doctor reports vary substantially; some recent LLM-based settings reach approximately 0.98 Accuracy and F1-score, while domain characteristics remain challenging.Early ML-based reports include Accuracy around 0.90 to 0.99 and F1-score above 0.90.Source: §6.1 · p. 28
  • Reported results show a shift toward representation learning, while Accuracy and F1-score diverge and traditional benchmarks show signs of saturation.The survey recommends complementary metrics and more realistic benchmarks, including LLM-generated reviews, adversarial attacks, multimodal content, and cross-platform evaluation.Source: §6.1 · p. 29

6.2. Comprehensive Comparison of Detection Paradigms

  • Paradigm choice involves trade-offs: ML favors small-data, lightweight settings; DL learns hierarchical patterns but needs more labels; PLMs balance semantic capability and feasibility; LLMs offer adaptive reasoning at higher deployment cost.The comparison spans learning paradigm, representation, annotation, adaptation, computation, interpretability, and generalization.Source: §6.2 · pp. 29, 31

6.3. Fusion-Oriented Practical Insights

  • Fusion combines evidence at feature, representation, graph, or decision levels; multimodal systems align review text, images, and behavior, but benchmarks and deployment remain limited.The survey identifies a lack of large-scale, realistic cross-platform multimodal benchmarks and notes computational, latency, and interpretability challenges.Source: §6.3 · pp. 31–32
  • LLMs enable fluent, controllable deceptive reviews while also supporting semantic detection and adversarial sample generation for detector robustness.The survey discusses generated hotel and product reviews, implicit deception signals, and the MAiDE-up benchmark [162].Source: §6.3 · pp. 31–32

7. Technological Challenges and Research Trends

The survey frames fake review detection as a changing, multi-source inference problem shaped by generated content, domain shift, and evidence reliability. It outlines research directions for robustness, adaptation, calibrated fusion, and accountable explanations.

7. Technological Challenges and Research Trends

  • The technical agenda addresses weaknesses in evidence reliability, benchmark validity, and multi-source fusion as detection moves into the LLM era.The survey presents unresolved challenges alongside research directions for addressing them.Source: §7 · p. 32

7.1. Challenges

  • Fluent AI-generated reviews and prompt-based style imitation can suppress detectable artifacts, requiring reasoning about semantic consistency, behavioral plausibility, and source reliability.Surface linguistic features alone are vulnerable to these changing attack strategies.Source: §7.1 · p. 32
  • PLM and LLM performance can degrade across platforms, domains, languages, and cultural contexts; architecture alone cannot solve transfer, which needs adaptation and explicit evaluation.Domain-adaptive pretraining can reduce part of the gap [224].Source: §7.1 · p. 32
  • Behavioral, temporal, visual, and graph evidence can be incomplete, noisy, private, or manipulated, so fusion must remain reliable when sources are missing or inconsistent.A multimodal approach combining review text, behavior, and ratings has been reported.Source: §7.1 · pp. 32–33
  • Moderation decisions need source-level explanations that show contributing evidence, conflict resolution, and uncertainty because false accusations can harm users and merchants.Text-only explanations may be insufficient when decisions rely on behavioral, graph, temporal, or multimodal evidence.Source: §7.1 · p. 33

7.2. Future Directions

  • Future work includes controlled generative augmentation, cross-domain adaptation, uncertainty-aware missing-source fusion, and verified explanations that mitigate hallucinated justifications.Suggested evaluation includes source-wise ablation, confidence calibration, and missing-modality stress tests.Source: §7.2 · p. 33
  • Benchmark auditing, source-wise ablation, and confidence calibration are immediate priorities, while continual defense and accountable human oversight need longer-term study.The broader direction is multi-source inference across textual, behavioral, structural, temporal, visual, and contextual evidence.Source: §7.2 · p. 33

8. Conclusion

The survey concludes that fake review detection is becoming a multi-source information-fusion problem rather than text classification alone, with future systems needing to account for evidence reliability and uncertainty.

8. Conclusion

  • Across methods from rule-based approaches to LLMs, the central challenge has shifted from surface anomalies toward semantic deception, manipulative intent, and evidence reliability.PLMs improve contextual semantic modeling but remain vulnerable to adversarial generation, domain shift, and limited interpretability.Source: §8 · p. 33
  • LLMs expand detection through prompt-driven reasoning, semantic representation learning, generative discriminative modeling, and data augmentation, while also enabling fluent fabricated reviews.Source: §8 · p. 33
  • The survey's main conclusion is that reliable detection requires integrating textual, behavioral, temporal, relational, multimodal, and external evidence while accounting for source reliability and uncertainty.The paper links this direction to reliable review moderation and preserving online reviews' consumer information value.Source: §8 · p. 33

CRediT authorship contribution statement

The authorship statement assigns data curation, investigation, and original-draft writing to Fanji Yang and Huiyao Chen.

CRediT authorship contribution statement

Related papers

Loading 2609.30292v1…