I think that the idea of analysing tone as a way to filter out scammers or bots is worthwhile, but I think it could also have the unintended consequence of picking up non-native English speakers, who might write in a slightly odd tone (but still be totally genuine). AI isn’t quite bias-free yet.
As long as the message tone captcha was combined with randomised captchas (for everyone - regardless of community standing/tone) then it would be fine. It could also be used for people who are copying and pasting the same message to multiple people (legit folks - but lazy couchers spraying hosts)