As the maxim goes, "The best way to learn is to teach". My vision here is that for any new topic a student learns, they (or the instructor) would be able to instantiate an AI agent with relevant preliminary knowledge, for the student to practice on. The student would try to teach the agent facts and/or how to perform basic tasks, and the agent, with some basic metacognition would be able to query the student regarding any unclear or conflicting points.
It definitely won't be anywhere near Turing Test level in 5 years, but I believe that by then we'll have something useful. And beyond that, I think there's real potential here, both for revolutionising education, and further down the line in terms of AGI.
This is slightly tangential, but this article from a few days ago strengthened my belief that we're getting closer - https://reiinakano.com/2019/11/12/solving-probability.html