The study claimed to be able to predict cannabis impairment using only the smartphone data with about 67% accuracy

As someone who studied data science and did a good bit of machine learning in that time: That's terrible and it means that the whole story is a big nothing burger. It's barely better than random guesses that match the distribution.