[jfim]

There are probably various metrics like language used to prompt the model, number of hours per day spent prompting, and many others.

There are quite a few other mitigations that could be done by providers that aren't mentioned in the article.