ChatGPT, an AI-powered conversationalist, has transformed how we work, learn, and write. Many use the tool since it’s easily accessible. However, problems have developed in distinguishing AI written content from human-written content, especially in academic and
professional writing settings. In this Cell article, Desaire et al. construct a machine-learning tool that identifies human academic scientific writing from ChatGPT writing with over 99% accuracy.
To do this, they trained a model using 64 articles published in the journal Science and 128 ChatGPT essays. After creating the model, they constructed two test sets containing 30 human writings and 60 ChatGPT writings. The model was then instructed to differentiate between AI-generated writing and human writing by examining paragraph complexity and sentence variety. The model was able to sort lines and words more effectively than an already publicly available AI detector tool, GPT-2 Output Detector.