Detecting Generated Text from ChatGPT (and other LLMs)



In this talk I discuss the implications of our research on the human detectability of generated text to educators grappling with the issue of students submitting generated homework assignments. I advise educators against using grammatical and fluency errors as evidence for generation and instead suggest they use factuality as their primary detection tool. Finally, I propose ways of writing assignments such as contrastive essays such that they are more difficult for generative models to complete.


This talk was given on February 10th 2023 to a virtual audience of University of Pennsylvania professors as part of the Critical Writing Seminar Panel for Educators.