As a frequent Redditor, this AI is going to be very witty.
They should limit it to top comments only, and for training, you might as well assume 90% of top comments are sarcastic/tongue in cheek. Or let a user dial the sarcasm/wittiness/seriousness as they want it, kind of like TARS from 'Interstellar'.
There are models for sentiment analysis (mostly using Twitter data, just enter sentiment analysis in Google Scholar) and you can use things like sentimentAPI: https://github.com/mikelynn2/sentimentAPI
They should limit it to top comments only, and for training, you might as well assume 90% of top comments are sarcastic/tongue in cheek. Or let a user dial the sarcasm/wittiness/seriousness as they want it, kind of like TARS from 'Interstellar'.