Detecting Large Language Models in Exam Essays

There is a widespread fear that large language models (LLMs) produce content that is indistinguishable from an original human work. Can students use LLM-based tools for their exam essays without being spotted? We performed experiments on a publicly available dataset that we produced, containing thirty answers provided by ChatGPT and thirty-six answers from university students, for each of six anonymized exam essay traces from a scientific methodology course for engineers. We applied term frequency-inverse document frequency (TF-IDF), a state-of-the-art machine learning algorithm for natural language processing, and the cosine similarity metric, in order to produce two LLM detectors. One is based on one-class support vector machine anomaly detection, and the other is based on multi-class random forest classification. The results show that it is possible to spot when LLMs are involved, provided that the source LLM is known. Education institutions can follow our guidelines to prevent cheating and improve education.