On the Linguistic Limitations of ChatGPT: An Experimental Case Study

Abstract

Artificial intelligence (AI) has become increasingly popular in recent years, due to advancements in computing power and data availability, as well as its broad range of applications in various industries. ChatGPT is a state-of-the-art language model developed by OpenAI, capable of generating human-like responses to text-based inputs. Its advanced natural language processing abilities have made it a valuable tool for a variety of applications, from customer service chatbots to language translation services. In general, similar to many other language models, ChatGPT uses an autoregression process to generate text. After training on massive data, ChatGPT predicts the likelihood of the next word in a sentence based on the previous words in the sentence. However, this autoregression working strategy of ChatGPT has caused some linguistic limitations. The first limitation is its tendency to generate text that exhibits a machine-patterned language. The second limitation is the generation of wrong invalid references. This paper conducts an experimental analysis of the linguistic limitations of ChatGPT caused by the word-by-word autoregression working strategy of ChatGPT. Three experiments will be conducted in this paper for proper evaluation: 1) Detecting texts written by ChatGPT. 2) Detecting text paraphrased by ChatGPT. 3) Detecting code written by ChatGPT. Extensive experiments involving a total number of 89,902 words and 552,381 characters have yielded compelling evidence that ChatGPT-generated text can be detected with a high percentage using AI techniques. Moreover, it was observed that AI-based detection methods did not consistently identify text paraphrased by ChatGPT due to the preservation of the original human structure within the paraphrased text. This also proves that there is a detectable difference between the human-based text structure and ChatGPT-based text structure.

Ähnliche Arbeiten