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.

We perform large-scale strong MIAs on pre-trained LLMs to clarify the extent of actual privacy risk MIAs pose in this setting.