I think there is a difference between comparing strings and so on, versus actually semantically processing the text. Spam/Phishing detection already feels invasive to me (although there is no way I would leave that to most users), and this is just one step further.