There is 2 things that are gonna help a lot, one is better classification, in that regard advances will help purge itself, just like our digestive system filters what we don't need models will do the same, e.g. scientific and factual models will have zero jokes in it's corpus, and you will know that the response you received came from such a humourless model, the second thing is just plain feedback, teachers will use AI and give it thumbs up or thumbs down and the machine will use such opinion over the opinion of most students, it will also favour the input by students with the most promise, those who win math championships and all that jazz.