Published Paper Details:

Large Language Models (LLMs) have made rapid strides in recent years, which has allowed them to excel at a wide range of activities including document completion and question responding. This has raised concerns about the unchecked usage of these models, which might lead to undesirable results like plagiarism, the creation of false news, spamming, etc. As LLMs must be used responsibly, accurate AI-generated content identification has become crucial. Several studies have attempted to solve this problem by including model signatures into text outputs or by watermarking texts with predetermined patterns. However, it has been observed that by a paraphrase attack, in which a light paraphraser is implemented on the LLM produced text. A range of AI detectors may be overwhelmed including those that employ watermarking methods, neural network-based detectors, and zero-shot classifiers. Furthermore, LLMs protected by watermarking methods are vulnerable to spoofing attacks, where a human adds covert watermarking signatures to human made text In this study, we examined the strengths and weaknesses of the watermarking and non- watermarking techniques for identification of text generated by AI. We looked at the soft watermarking technique for AI generated text and non-watermarked AI generated text and observed how it was vulnerable to spoofing and paraphrase attacks. This study reveals that the AI generated text detectors are unreliable under real-world conditions.