Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection

Glimpse enables white-box methods to utilize proprietary models for zero-shot LLM-generated text detection, overcoming limitations of current techniques. Experiments demonstrate that Glimpse with Fast-DetectGPT and GPT-3.5 achieves an average AUROC of about 0.95 across five latest source models, improving scores by 51% relative to open-source baselines. This suggests advanced LLMs can effectively detect their own outputs, potentially making them their own best shield.