I was reading something about how google's new chatbot fails the turing test when you try to elicit machine behaviour from it. Could one train a chatbot that passes the turing test by concurrently training two nets, one that is the chatbot and another that detects whether it's a chatbot or a real conversation? I inagine that this should work theoretically, but I'm by no means an expert.
> and another that detects whether it's a chatbot or a real conversation?<p>Problem is: How would it do that? In order to for a machine to accurately decide whether its conversation partner is a human or another machine, it would require a model that already accurately understands the difference.<p>So it's basically a hen-and-egg problem; In order to make the Discriminator for a GAN that should pass the Turing test, I need a machine that could already pass the Turing Test.<p>Sure, we could feed it billions of human conversations, and train it to recognize some sort of difference, and that could result in a Generator which somewhat mimics the way human conversation works, but that's not what the Turing Test evaluates...when I ask something that is nonsensical like "what does the magnetopause smell like", and the system simply simulates a human speech pattern along the lines of "i don't know, I never tried smelling it", instead of showing <i>actual understanding about the universe</i> that tells it that the question is absurd, it may fool the Discriminator-Algorithm of the GAN, but its not going to convince humans for long.