Trust at risk: Detecting misinformation in LLM-generated product reviews and its implications for consumer behavior and platform governance

The rapid adoption of large language models (LLMs) has intensified the risk of AI-generated fake reviews that distort consumer perception and erode trust in digital marketplaces. This study addresses this challenge by constructing a multi-domain synthetic dataset of hotel, restaurant, and product reviews. Synthetic reviews were generated through three Controllable Misinformation Generation (CMG) strategies: paraphrasing, rewriting, and open-ended prompting using ChatGPT-3.5 and GPT-4 with human validation for semantic consistency and fluency. Detection performance was evaluated across zero-shot instruction-tuned LLMs (LLaMA2-7B, LLaMA2-13B, ChatGPT-3.5, and Gemini-2.5-Flash) and supervised transformers (RoBERTa-base and DeBERTa-v3-base). Results show that open-ended generations are substantially harder to identify, with accuracy drops exceeding 25% compared to constrained styles. DeBERTa-v3 achieved state-of-the-art performance, reaching 96%–98% accuracy/F1 on paraphrased and rewritten reviews (restaurants/Amazon) and consistently outperforming RoBERTa, while zero-shot detectors achieved success rates below 45%. We further observe that larger model size does not guarantee better zero-shot detection; for instance, LLaMA2-7B occasionally outperforms 13B under comparable prompting, underscoring sensitivity to prompt design and sampling settings. These findings underscore the critical limitations of instruction-tuned LLMs in authenticity detection, emphasizing the urgent need for hybrid, domain-adaptive moderation pipelines that integrate robust supervised detectors with flexible LLM-based modules, a focus that is pivotal for effective platform governance and evolving regulations on fake reviews. Future work should expand this research by developing multilingual, cross-domain datasets, improving adversarial robustness through stress testing, and deploying hybrid detection systems enriched with behavioral and metadata signals.