It's a question of generalizability. Today's "AI" algorithms are intensely overfitted to their problem domains, even if they generalize well within those domains.