How might an AI explain itself?