Deep learning is great for learning correlations in large amounts of data, but I doubt its efficacy for learning the kinds of causation and constraints in systems that depend on rules defined by people. An example is learning the rules of chess just by observing the moves of pieces in a large number of games. A deep learning AI might get fairly far at that, but could it learn infrequent moves like en passant pawn capture or pawn promotion to pieces other than queens? Of course it could only learn those moves if it saw such moves occur. And could it learn that some moves are never permitted, such as castling after the king or rook involved has already been moved or when the king would be moving through check? How would it know that a bishop can never move to a square of the opposite color from the square it is currently on, and that, unlike a knight, it can never jump over another piece?

I think learning of some kinds of human rule-based systems — games, laws, grammar rules, arcane but rigid social conventions and the like — necessarily require some explicit teaching or at least are greatly facilitated by explicit teaching.