It’s plainly false that they cannot rule out whether their de-identified data was used in training their model. Just that they haven’t ruled it out.