Research Explores Making AI Text Indistinguishable from Human Writing.

The rapid advancement of large language models (LLMs) has made it increasingly difficult to differentiate between AI-generated and human-written text. This challenge is further compounded by the emergence of paraphrasing tools designed to make machine output appear more natural and human-like. This paper delves into the strategic use of human writing samples to guide the paraphrasing process. The study proposes a multi-sample setting where both human and machine responses to the same prompts are available. It shows that through repeated paraphrasing, the distribution of machine-generated text can be systematically shifted towards the empirical distribution of human writing. This convergence occurs under simple mixing and stability conditions. The research provides explicit convergence rates and characterizes how the necessary number of human samples and paraphrasing rounds scale with the desired level of indistinguishability. This work highlights the growing difficulty in detecting AI-generated content and the potential for sophisticated manipulation, raising important questions about content authenticity and detection methods.