An Empirical Study to Evaluate AIGC Detectors on Code Content

Artificial Intelligence Generated Content (AIGC) has garnered considerable attention for its impressive performance, with Large Language Models (LLMs), like ChatGPT, emerging as a leading AIGC model that produces high-quality responses across various applications, including software development and maintenance. Despite its potential, the misuse of LLMs, especially in security and safetycritical domains, such as academic integrity and answering questions on Stack Overflow, poses significant concerns. Numerous AIGC detectors have been developed and evaluated on natural language data. However, their performance on code-related content generated by LLMs remains unexplored. To fill this gap, in this paper, we present an empirical study evaluating existing AIGC detectors in the software domain. We select three state-of-the-art LLMs, i.e., GPT-3.5, WizardCoder and CodeLlama, for machine-content generation. We further created a comprehensive dataset including 2.23M samples comprising coderelated content for each model, encompassing popular software activities like Q&A (150K), code summarization (1M), and code generation (1.1M). We evaluated thirteen AIGC detectors, comprising six commercial and seven open-source solutions, assessing their performance on this dataset. Our results indicate that AIGC detectors perform less on code-related data than natural language data. Fine-tuning can enhance detector performance, especially for content within the same domain; but generalization remains a challenge. CCS CONCEPTS • General and reference → Empirical studies; • Security and privacy → Social aspects of security and privacy; • Computing methodologies → Natural language generation.