Answers are grounded in this paper and cite their sources.
Large language models have made fluent fake reviews cheap to produce, eroding the reliability of the linguistic surface that e-commerce platforms have long used to police authenticity. We propose to detect inauthentic reviews not from text in isolation but from their consistency across the multiple relational contexts a review inhabits: its product, its rating, its writing-effort window, and its semantic neighborhood. We model these contexts as a multi-relational hypergraph and equip it with a cellular-sheaf operator whose per-relation consensus residuals quantify how far a review departs from the local agreement of its peers. The resulting Multi-Relational Sheaf-Hypergraph Network (MR-SHN) produces an interpretable, per-relation authenticity profile suitable for trust-and-safety review. On a unified corpus of 46,154 reviews spanning three domains and four authorship conditions, in-distribution detection is essentially saturated (F1 up to 0.980), yet all detectors collapse under domain shift, with AUC falling to 0.48. We document a marked asymmetry: detectors trained on human-written deception transfer to GPT-4-Turbo reviews (AUC0.88), while the reverse fails. A training-free residual score separates authentic from modern AI reviews at AUC0.989, and the model transfers to six unseen generator families at AUC0.83. We report robustness limits candidly and discuss implications for platform governance.
No takes yet. Share an insight, caveat, or question.
Balcı et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: