In a recent Chronicle of Higher Ed article, Shreeharsh Kelkar (University of California at Berkeley) argues that watermarking offers a solution to the issue of students cheating with large language models (LLM). Kelkar explains how watermarking—where a detectable pattern is left in machine-generated text—could help identify this text without impairing AI performance. Kelkar references two methods of watermarking: LLM text watermarking by generative AI companies and in-context watermarking that would use a learning management system program to alter assignment prompts so that any ensuing LLM output would have a watermark. Kelkar concludes with a discussion of the obstacles to implementing watermarks and calls for stakeholders to come together and find a way to reliably detect these outputs.
