We develop tests for proving that someone is producing text using a particular language model by correlating the text with the order of examples used to train the model.

A technical report detailing a NeurIPS competition testing the robustness of popular watermarks and practical guidance for future development.