[Paper Review] Detecting ChatGPT: A Survey of the State of Detecting ChatGPT-Generated Text

Mahdi Dhaini, Wessel Poelman+1 more|arXiv (Cornell University)|Sep 14, 2023

Artificial Intelligence in Healthcare and Education9 citations

TL;DR

A comprehensive survey of datasets and methods for detecting ChatGPT-generated text, focusing on academic works, datasets, and qualitative insights, with discussion on robustness and multilinguality.

ABSTRACT

While recent advancements in the capabilities and widespread accessibility of generative language models, such as ChatGPT (OpenAI, 2022), have brought about various benefits by generating fluent human-like text, the task of distinguishing between human- and large language model (LLM) generated text has emerged as a crucial problem. These models can potentially deceive by generating artificial text that appears to be human-generated. This issue is particularly significant in domains such as law, education, and science, where ensuring the integrity of text is of the utmost importance. This survey provides an overview of the current approaches employed to differentiate between texts generated by humans and ChatGPT. We present an account of the different datasets constructed for detecting ChatGPT-generated text, the various methods utilized, what qualitative analyses into the characteristics of human versus ChatGPT-generated text have been performed, and finally, summarize our findings into general insights

Motivation & Objective

• Motivate the need to distinguish human- from ChatGPT-generated text across domains (law, education, science).
• Review datasets and methods developed specifically for ChatGPT detection.
• Summarize qualitative insights on linguistic and stylistic differences between human and ChatGPT text.
• Highlight challenges, limitations, and directions for future research in detection.
• Provide a consolidated view to guide researchers in dataset and method selection.

Proposed method

• Review and categorize detection approaches (classifier-based, zero-shot, perplexity-based, and explainability-driven).
• Compile and summarize datasets created for ChatGPT detection (domain, language, public availability, and out-of-domain variants).
• Compare methods based on performance, robustness to adversarial prompts, and multilingual capabilities.
• Extract qualitative linguistic characteristics distinguishing human vs. ChatGPT text across domains.
• Discuss explainability tools (e.g., SHAP) used to analyze detectors’ decisions.
• Synthesize general insights and outline open challenges and future directions.

Experimental results

Research questions

• RQ1 What datasets exist for detecting ChatGPT-generated text, and are they publicly accessible?
• RQ2 What detection methods have been proposed specifically for ChatGPT-generated text, and how do they differ across languages and domains?
• RQ3 How do linguistic and stylistic features differentiate human and ChatGPT writing across domains?
• RQ4 What are the key challenges and robustness concerns for ChatGPT-detection methods (adversarial attacks, out-of-domain text, multilinguality)?

Key findings

• Multiple datasets exist across domains (education, science, medicine, general QA).
• RoBERTa- or transformer-based detectors often outperform perplexity-based baselines, with robustness to some adversarial variants.
• Language models trained on English data generally outperform multilingual models in detection tasks.
• Explainability methods (e.g., SHAP) help interpret detector decisions and identify salient features.
• Detection performance degrades on out-of-domain data and with adversarial alterations (misspellings, homoglyphs).
• Shorter text lengths reduce detection reliability; longer, full-context texts yield better results.