Answers are grounded in this paper and cite their sources.
Generative language models can produce large volumes of fluent, product-specific online reviews, while many existing detectors are evaluated on a single generator or rely on comparatively large models. This empirical, systems-oriented study examines whether an established parameter-efficient pipeline can distinguish Original Reviews ( or ) from Computer-Generated Reviews ( cg ) efficiently and retain useful performance when the generator changes. The labels indicate experimental text provenance and do not imply truthfulness, deception, or fraudulent intent. A 4-bit Qwen2.5-0.5B sequence classifier was adapted with rank-64 rank-stabilized quantized low-rank adaptation (Rs-QLoRA; α = 128), updating 35.2 million parameters (6.65% of the loaded model) within approximately 2.05 GiB of peak PyTorch-allocated training memory. In addition to a GPT-2-era corpus, contemporary review datasets were constructed with LFM2.5-1.2B-Instruct, DeepSeek-V4-Pro, and GLM-5.2. Because these datasets share a source-review pool, they were also evaluated under a source-grouped partition that assigns each original review and all of its generated derivatives to a single partition. Across three training seeds, matched accuracy was 98.497–99.776% under row-level partitions and 99.552–99.758% under source grouping. Transfer was weaker and directional: with source reviews isolated, off-diagonal cg recall among the contemporary generators averaged 73.351% (range 31.425–97.384%), while every GPT-2/contemporary direction in the row-level four-domain matrix remained at or below 0.5%. Pooled training on two contemporary generators detected 95.240–97.149% of cg reviews from the held-out generator at false-positive rates of 0.448–0.742%, matching or exceeding the best single training generator. Rs-QLoRA and standard QLoRA reached similar accuracy, but Rs-QLoRA stopped after fewer optimization steps in all six paired runs. With the rank-stabilized multiplier held fixed, ranks 16–64 differed by at most 0.101 percentage points. At an assumed 1% cg prevalence, expected alert precision at the default threshold was 57.8–68.7% for held-out generators. Applied without retraining to three independently collected corpora (MAiDE-up, DetectAIRev, and MAGE-Yelp), the same detectors recovered only 0.013–35.711% of cg reviews at false-positive rates of 0.113–0.941%, with ROC-AUC of 45.426–80.514%. A compact detector can be highly competitive on matched distributions and transfers among generators that share one generation framework, but neither matched accuracy nor within-framework transfer establishes generator-independent detection.
No takes yet. Share an insight, caveat, or question.
Rhrissi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: