Screw the Turing test — chatbots don’t need to act human

The minute you mention chatbot technology, someone will inevitably bring up the Turing test. Created by Alan Turing in 1950, the test judges a machine’s capability to demonstrate intelligent behavior that’s indistinguishable from speaking to a human. There’s even an annual competition that started in 1990 called the Loebner Prize that judges chatbots on how humanlike they can be. In each round of the competition, a human judge concurrently holds a chat-based conversation with both a chatbot and human being via a desktop computer. Based upon the chat conversation, the judge must try to figure out which is the chatbot.

While this competition is great to push the technology of chatbots and AI forward, it’s kind of a trap. This technology isn’t about fooling people — it’s about making people’s lives better. Everyone who logs into our system knows that our health care chatbot Joy is not a person: Her avatar has blue skin, for starters. Instead of trying to fool our members, we go out of our way to show that she’s a chatbot. She is supported by our human concierge staff, and we make it very clear when we do a handoff. Our staff has real pictures, and we even make slight changes to the chat window when they type.

[aditude-amp id="flyingcarpet" targeting='{"env":"staging","page_type":"article","post_id":2166358,"post_type":"guest","post_chan":"none","tags":null,"ai":true,"category":"none","all_categories":"ai,bots,","session":"D"}']

We want to make a clear distinction in who’s attending to you because we are dealing with our members’ health. Our members are very comfortable chatting with Joy, but there are some things that can only be done by a person. We are big believers in the human-augmented chatbot approach, especially in health care.

Joy does use natural language processing (NLP), and we are improving the technology every day, but we actually script most interactions. Health care is complex, and people need guidance in asking the right questions. Sometimes people know they need something but aren’t even sure what they are looking for. While NLP is incredibly powerful technology (and it’s improving our Net Promoter Score), our members still need scripted interactions to guide them to the best care. Writing chatbot scripts can be a little tough, but it’s vital for more complex industries.

I would think of chatbot technology in the same light as a trip to Disneyland. When interacting with a fictional character, people are willing to have a suspension of disbelief and enjoy the interaction. In our user testing, I’ve seen people interact with Joy. When they get a response with an emoji, I’ve seen the smiles. People know it’s not real and don’t care that it wouldn’t pass a Turing test — especially when Joy saves them money. Just think of the character Baymax from the movie Big Hero Six: He doesn’t sound or look human, but that doesn’t prevent you from empathizing with him. Would Baymax pass a Turing test? No, but that’s not what matters.

I’m not saying that there isn’t value in making chatbots more humanlike. It’s amazing to see the improvements that people are making every day. I’m only saying that the end goal of a chatbot isn’t to be humanlike; it’s to help people and to deliver results.

I think most people’s frustration in dealing with a bad chatbot is the underlying NLP technology. Take Siri as an example. I don’t care if she never speaks to me again because I’m usually looking down at the screen of my iPhone when I’m interacting with her.  All I want is for her to deliver me the results I’m looking for and to be able to have more context in delivering those results. I only get frustrated with Siri when she doesn’t understand what I’m talking about. I’ve never thought that her responses needed to be more humanlike. Yeah, the parlor tricks like telling jokes can be fun, but when was the last time you really interacted with her in that way? I never need to hear another bad joke from Siri in my life.

The same goes for Alexa. I have one at home and love it. It’s an amazing device and works extremely well for the tasks that are most important to me within a home setting using voice. Alexa would never pass a Turing test, but she delivers results, and that’s what matters to me. It’s nice when you’re in the kitchen cooking and you can ask her to turn on the lights, put on a timer, and play a little music. It’s the convenience she delivers, not her personality, that makes me love Alexa. So for now, don’t concern yourself with the Turing test. Enjoy the technology for the results it can deliver.