Venue
International Conference on Machine Learning
Domain
natural language processing
The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for Detect-GPT. See ericmitchell.ai/detectgpt for code, data, and other project information.
This paper presents DetectGPT, a zero-shot method for detecting text generated by large language models (LLMs) based on the curvature of the model's log probability function. It leverages the observation that model-generated text tends to occupy regions of negative curvature in the probability landscape, making it detectable without the need for training a separate classifier or collecting labeled data. The authors argue that their method improves detection accuracy compared to existing zero-shot methods. They conduct empirical tests across multiple datasets, demonstrating that DetectGPT outperforms other methods in identifying generated text across a range of models and conditions, including the detection of misinformation. The paper also discusses the implications of model generation in fields like education and journalism.
This paper employs the following methods:
- GPT-3
- GPT-2
- GPT-Neo
- GPT-NeoX
- T5
The following datasets were used in this research:
- XSum
- SQuAD
- Reddit WritingPrompts
- WMT16
- PubMedQA
- 0.95 AUROC for DetectGPT
- Improved detection from 0.81 to 0.95 AUROC
The authors identified the following limitations:
- Number of GPUs: None specified
- GPU Type: None specified
machine-generated text detection
zero-shot detection
probability curvature
log probability function