Why’s it so hard to teach robots to talk?

Natural language processing using real-world, data-based training methods can teach chatbots to talk and improve the effectiveness of modern applications of artificial intelligence, but how can we be sure it’s ethical?