A Ph.D. student at McGill University is coaching artificial intelligence programs to pass the Turing test.
The Turing test is a test that evaluates a machine or program's intelligence. The test will determine if the machine's intelligence is equivalent to or even indistinguishable from that of humans. With this in mind, Ph.D. student Ryan Lowe is developing a chatbot that can pass as human to the people interacting with it.
It's often obvious to us when we're interacting with a bot and not a human. Lowe aims to basically fool us into thinking that we're interacting with an actual human being.
What, however, would this AI be for? For one thing, it can be a far better virtual assistant than the ones we interact with now. If you've ever had to speak to tech support, you'd now how frustrating it is to interact with automated programs. Thus, chatbots that can successfully mimic human interaction may be something to look forward to.
How do we know if an AI program passes the Turing test? Though there is some disagreement between scientists, most agree that if a program can fool most people that interacts with it, then it passes. Amazon, for example, uses human testers to see how well its voice-operated personal assistant Alexa interacts with users. Then again, Amazon is a huge company. Smaller companies may not have the resources to test their programs with large groups of people.
So what was Lowe's answer to this problem? He thought to use an AI system to evaluate how human-like a chatbot's dialogue sounds, basically using AI to evaluate AI. He then took 1000 Twitter conversations and asked human volunteers to add a reply to these conversations. Then, he let chatbots add a reply to these conversations as well. Another group of human volunteers then rated the chatbots' replies.
Lowe used these ratings to teach his system how to distinguish between replies that passed as human and the ones that don't. This enabled the system to make the same judgments that the human evaluators did.
Of course, a chatbot that can give human-like replies can be a double-edged sword. There are times when we'd rather speak to a human than a machine, and other times when we'd rather speak to a human. A human-like support chatbot can have the benefits both of human and machine support, but it can also have the disadvantages. For example, the chatbot may give very human replies that can be frustrating and unhelpful.
This is something that Lowe aims to tweak in the system. After all, if you'll be using a chatbot that can pass the Turing test, while leave the disadvantages in? Since other researchers can also use Lowe's evaluator system for their own projects, human-like chatbots may soon be available to us.
Get weekly science updates in your inbox!