AI-generated plagiarism is difficult to detect, but it is possible

According to Pudasaini et al. (2024), AI-generated plagiarism is placed at the most difficult level to detect (Figure 3). The easiest to detect is character-preserving plagiarism, such as copy and paste. The next level is syntax-preserving plagiarism, which involves synonym substitution. This is followed by semantics-preserving plagiarism, such as translation and paraphrasing. The next most difficult type is idea-preserving plagiarism, which includes structural plagiarism and the use of others' ideas and concepts. Ghostwriting, or hiring someone else to write, comes next. The most difficult to detect is AI-generated plagiarism, which involves using AI to generate text and even employing AI to bypass plagiarism detection systems.

Plagiarism Detection Difficulty Level (Pudasaini et al., 2024)

According to Pudasaini et al. (2024), LLM-generated text may be detected not only through traditional plagiarism detection methods but also using LLM-generated text detection techniques. These techniques include approaches such as watermarking, zero-shot detection, and classifier-based training. 

In our previous article titled "Does ChatGPT Plagiarize? The Plagiarism Checker Results Show a High Percentage," we showed that generated text can also be detected by traditional plagiarism checkers, not just using LLM-generated text detection techniques. Based on our findings, generated text from AI combines text from a variety of sources. It is detected by our plagiarism checker.

Reference

Pudasaini, S., Miralles-Pechuán, L., Lillis, D. et al. Survey on AI-Generated Plagiarism Detection: The Impact of Large Language Models on Academic Integrity. J Acad Ethics (2024). https://doi.org/10.1007/s10805-024-09576-x