Testing the Limits: Evaluating AI Detectors’ Accuracy and the Impact of Obfuscation Techniques on AI-Generated Text

The rise of Artificial Intelligence (AI)-generated text has led to the development of numerous detection tools to distinguish between human and machine-authored content.However, the effectiveness of these tools, especially against manipulated texts, remains uncertain.This study evaluates nine widely used AI detection tools-Turnitin, ZeroGPT, Detecting-AI.com,GPTZero, QuillBot, Grammarly, Sapling, Copyleaks, and Originality.ai-usingtexts from four large language models-ChatGPT, DeepSeek, Gemini, and Grok-as well as human-written samples.Initial findings indicate that commercial tools, such as Copyleaks and Originality.ai,achieved near-perfect detection rates, while free tools, including Grammarly and QuillBot, performed less reliably, with some as low as 63.0%.On the other hand, paraphrasing and Non-Native English Speakers (NNES)-style rewriting techniques reduced detection accuracy across most detectors.Turnitin dropped to 45.7%, while Grammarly fell to 19.0% in some cases.Only Copyleaks, GPTZero, and Sapling maintained strong performance under obfuscation.The study highlights three issues: inconsistent detector performance, the impact of obfuscation, and ethical risks, including bias and false positives.The study suggests that while some detectors offer robust baseline performance, combining them with pedagogical strategies and policies is essential to uphold academic integrity.

Перевод пока недоступен