We build systems to organize who we are (urbandictionary) but hate it when the systems use that information tell us who we are (watson).
It feels so much like the emperor isn't wearing any clothes.
Perhaps an appropriate response would be for the computer to measure the tension in the human voice response to it's queries and optimize for lower tension.
So it can pick three words: Bullshit -> 80% confident; Sham -> 70% confident; Fallacy -> 50% confident;
Within limits, it will pick the less optimal word and measure the tension in the response and find a way of influencing confidence based on responses.
Think multi-armed bandit problem but with social situations. I mean, to be honest, isn't that what we all did when we were in middle school? We used as many bad words as possible measuring the response we got from others? None of us were born with a binary understanding of when to use certain words it was more trial and error.