33-46% of MTurk crowd workers used LLMs when completing a task

Researchers Veniamin Veselovsky, Manoel Horta Ribeiro and Robert West report that many crowd workers likely used large language models when completing text tasks on Amazon Mechanical Turk. In a case study, they reran an abstract summarization task from the literature and, combining keystroke detection with synthetic text classification, estimated that 33-46% of workers used LLMs. The authors note that generalization to less LLM-friendly tasks is unclear, and they call for platforms, researchers and workers to find ways to keep human data human.