GPT-Sentinel: Distinguishing Human and ChatGPT Generated Content

TL;DR

The paper tackles the problem of distinguishing human-authored text from ChatGPT-generated content by creating the OpenGPTText dataset through paraphrase-based generation and training two detectors, RoBERTa-Sentinel and T5-Sentinel. It compares supervised fine-tuning approaches against traditional and zero-shot methods, demonstrates high accuracy (>97%) on test data, and conducts a multifaceted interpretability study (PCA, integrated gradients, and t-SNE) to reveal the features driving GPT-like predictions. The work reveals that while both models perform well on the OpenGPTText-derived data, transfer to GPT-2 outputs remains challenging, and the T5-based approach generally offers more robust performance with higher confidence. The dataset release plan and interpretability analyses enable practical deployment and future improvements across languages and contexts.

Abstract

This paper presents a novel approach for detecting ChatGPT-generated vs. human-written text using language models. To this end, we first collected and released a pre-processed dataset named OpenGPTText, which consists of rephrased content generated using ChatGPT. We then designed, implemented, and trained two different models for text classification, using Robustly Optimized BERT Pretraining Approach (RoBERTa) and Text-to-Text Transfer Transformer (T5), respectively. Our models achieved remarkable results, with an accuracy of over 97% on the test dataset, as evaluated through various metrics. Furthermore, we conducted an interpretability study to showcase our model's ability to extract and differentiate key features between human-written and ChatGPT-generated text. Our findings provide important insights into the effective use of language models to detect generated text.

Figures (15)

• Figure 1: PCA of hidden state distribution of RoBERTa-Sentinel model on OpenWebText (Left) and OpenGPTText (Right) before and after cleaning. Note that the cleaning process affected the distribution of OpenWebText significantly.
• Figure 2: The depicted figure illustrates the RoBERTa-Sentinel architecture, wherein the dashed line connecting RoBERTa-Base and MLP module indicates the non-propagation of gradient back to the former.
• Figure 3: Architecture for T5-Sentinel. After input the entire token sequence, we provide the T5-Decoder with a <PAD> token and predict if the input text is by human or generated based on the probability of specific word "Positive" and "Negative" in the output word probability distribution.
• Figure 4: ROC Curves for models across different data sets. OpenGPTText-Final (Left), OpenGPTText (Middle), and GPT2-Output (Right)
• Figure 5: ROC Curves for same model under different data sets T5-Sentinel (Left) and RoBERTa-Sentinel (Right). Note that the performance of RoBERTa-Sentinel significantly deteriorates when transfer to original version of OpenGPTText while T5-Sentinel does not.