Google tries to make Duplex AI phone calls less creepy

So what you gonna hear is the google assistant actually calling a [UNKNOWN] to schedule appointment for you. Let listen. [SOUND] Hello, how can I help you? Hi, I'm calling to book a women's haricut for client, I'm looking for something on May 3rd. Sure, give me one second. Mm-hmm. [LAUGH] [MUSIC] Do you have anything between 10 am and 12pm? Depending on what service she would like. What service is she looking for? Just a woman's haircut, for now. Okay, we have a 10 o'clock 10 AM is fine. El Cocotero, how may I help you? Hi, I'm the Google Assistant calling to make a reservation for a client. This automated call will be recorded. Can I book a table for Tuesday, the 12th? Okay, cool. We've always been thinking about how should the system kind of introduce itself. And the way that we think about it is really having four key elements. First is, starting by saying what the purpose of the call is, because for the business, that's really the important thing, calling for a reservation or what are you calling for? Second is, disclosing that this call is from Google. Third, that it's from an automated system. So not just a human operator at Google, but an automated system. And then fourth in states that require it, we will disclose that we're recording the call. Great, and when did you say they wanted to come in? Tuesday at seven pm? Okay, let me check. Making those systems sound and feel more natural you have a higher success rate of completing the call successfully for kind of the business side and the customer side. And so as we kind of study that and look for ways to make it natural, we came upon this idea that linguists call speech disfluencies ums, and ahs that we all use every day in conversation. As linguists have studied this they've found that those words or non words actually play a crucial role in helping the conversation to continue. And so they play some really important purposes in our system to help both sides of the conversation kind of know that the other side's still listening, or introduce things like corrections And clarifications. Hey Google, book a table for two at [UNKNOWN] on Tuesday at seven. We're testing, we're gonna go kinda slow. Take it step at a time. So we're starting by testing with a group of trusted testers, a fairly small group and a set of partner businesses that we work with. Really to get feedback both from those end users and from the businesses in how they respond to this kind of technology. All right, I'll call to book under your name and phone number, and I'll update you in the next 15 minutes. Is that okay? Perfect, thanks. A lot of the concerns that we heard about Duplex and good concerns as we announced an IO. But some of them were kind of around wow has Google created a general AI that can have any conversation and act like a person, and how will people know what it is. And we wanna make clear that it's very far from what we've actually done. What we actually done is train very deep models that know how to do very narrow task. Holiday hours, restaurant reservations, booking a haircut appointment, those are really the three things that, so if you think duplex is like a human, it's a very interesting human that can only do those three things. I don't have 7 but we could do 8. Where the system doesn't know what the right next step is. And so within the calls what you'll hear is, the system will try a couple times usually to kinda bring it back to what it understands. So if you're talking system and you say, well, Something crazy. You're Google, well what's the weather tomorrow? What you'll hear the assistant do is say something like, I'm not sure I understand but could I get a table for four on Tuesday? Right, so it kinda try to bring it back. If that doesn't work, the system is able to signal a human operator And that human operator can either come in and just step in on the call. So what you'll hear is someone saying, sorry, it sounds like our automated system was having trouble completing the reservation. This is Dave, I'm a human, and let's continue the conversation. Or, alternatively, the automation can simply bow out and say, I'm not sure we're Getting through this, or let me have someone call you back and then have the human call back. Okay, we'll see you on Tuesday, thank you. Okay, awesome. Thanks a lot.