Complexity and Similarity Analysis of ChatGPT-Generated Text

Abstract

ChatGPT, introduced by OpenAI in November 2022, is a remarkable development in this field. It is a Large Language Model, one of the largest models, trained with 175 billion parameters taken from the web. ChatGPT's ability to produce text closely resembling human writing poses a significant challenge in distinguishing between machine-generated and human-generated content. Given its ability to generate almost any type of text, students and researchers may be tempted to use ChatGPT to write papers. Moreover, students may employ it to complete assignments, such as essay writing, answering questions, and even programming tasks, as the model can perform these tasks without raising suspicion. The students are not stimulating their brains to complete the tasks during the learning process, which raises the problem of how a professor from any educational level can detect if any task was completed using ChatGPT. We know that ChatGPT was trained using billions of data from the web, so one way to detect if a text was created with ChatGPT is that it will have a high similarity index and the technical level of text. For that reason, this study presents an analysis of similarity of 100 ChatGPT's responses to three questions with other web content and evaluating their complexity using readability, comprehensibility, and perspicuity. We also analyzed 21 texts generated by humans that answered the same questions as ChatGPT to evaluate the same metrics that were mentioned before.

Ähnliche Arbeiten