On the Detectability of ChatGPT Content: Benchmarking, Methodology, and Evaluation through the Lens of Academic Writing

TL;DR

The paper addresses the challenge of detecting AI-generated content in academic writing by introducing GPABench2, a large, cross-disciplinary benchmark of 2.385M abstracts across CS, physics, and HSS, and CheckGPT, a model-agnostic detector built on RoBERTa-LSTM with strong transferability. It demonstrates that humans struggle to differentiate GPT-generated from human-written abstracts and that existing detectors are often inadequate in this niche, motivating the development of CheckGPT. Extensive experiments show CheckGPT achieving near-perfect accuracy across tasks, disciplines, prompts, and even unseen LLMs, with robust performance under advanced prompts and sanitization attacks. The work provides actionable insights into domain transfer, prompt design, and the evolving use of AI in scholarly writing, and it releases open resources to support responsible AI governance in academia.

Abstract

With ChatGPT under the spotlight, utilizing large language models (LLMs) to assist academic writing has drawn a significant amount of debate in the community. In this paper, we aim to present a comprehensive study of the detectability of ChatGPT-generated content within the academic literature, particularly focusing on the abstracts of scientific papers, to offer holistic support for the future development of LLM applications and policies in academia. Specifically, we first present GPABench2, a benchmarking dataset of over 2.8 million comparative samples of human-written, GPT-written, GPT-completed, and GPT-polished abstracts of scientific writing in computer science, physics, and humanities and social sciences. Second, we explore the methodology for detecting ChatGPT content. We start by examining the unsatisfactory performance of existing ChatGPT detecting tools and the challenges faced by human evaluators (including more than 240 researchers or students). We then test the hand-crafted linguistic features models as a baseline and develop a deep neural framework named CheckGPT to better capture the subtle and deep semantic and linguistic patterns in ChatGPT written literature. Last, we conduct comprehensive experiments to validate the proposed CheckGPT framework in each benchmarking task over different disciplines. To evaluate the detectability of ChatGPT content, we conduct extensive experiments on the transferability, prompt engineering, and robustness of CheckGPT.

Figures (13)

• Figure 1: The architecture of the CheckGPT model.
• Figure 2: Training losses of the task-specific and discipline-specific classifiers.
• Figure 3: Feature space distribution of human-written (green) and GPT-generated (red) abstracts.
• Figure 4: CheckGPT's transferability across disciplines and tasks: (a) without fine-tuning, (b): tuned with 5% data from the train set. 1C: Task 1 GPT-WRI+CS; 2P: Task 2 GPT-CPL+PHX; 3H: GPT-POL+HSS.
• Figure 5: TPR on ChatLog-HC3 with direct validation, fine-tuning, and full re-train.