Doesn't that come down to allowing it to directly regurgitate training data? Surely it's seen dozens of such solutions.