Google's New Phone Assistant Sounds Eerily Human. Is That a Good Thing?

In an effort to quell concerns about its newest artificial intelligence technology, Google invited a handful of reporters to a quaint hummus shop in downtown Mountain View to test it out.

The virtual assistant, Duplex, made a huge splash during the Google I/O developer conference in May. It sounded so much like a human that many people thought it was fake. Critics immediately began talking about the dangerous potential of having a robot sound like a real person.

“I think one thing that’s really important to understand about the system, and how it’s built, is what makes it work is that it only is able to do very narrow tasks,” explains Scott Huffman, vice president of engineering for the Google Assistant.

To showcase the scope of this technology, myself and a handful of reporters were each given about a minute and a half to demo the system. I took off my reporter hat and turned into a maitre d’ at Oren’s Hummus Shop, standing at the front of the restaurant answering the phone.

“Hi, I’m calling to make a reservation,” the Google assistant said, sounding like a 20-something woman, perhaps someone I might be friends with.

“Hi! How can I help you?” I asked.

“I’m Google’s automated booking service,” the robot said, “So I’ll record the call. Um, can I book a table for Saturday?”

I tried a few tricks to trip her up. At one point I talked over her and asked if she could repeat what she said. There was a pause, and then the robot replied, “Um, I’d like to make a reservation for four people please.” The robot was friendly, direct and human-sounding, even though she’s not.

What’s with the “um”? Huffman says when his team initially tested the reservation system out with a robotic-sounding voice, it didn’t work. “Restaurants were more likely to hang up on it,” says Huffman.

A human-sounding voice, he says, with all of the “ums” and “ahs” actually gives a signal to the reservation taker that the call is not a scam or fake. Huffman says it makes the task of booking a reservation easier.

But critics like San Jose State professor Ahmed Banafa worry that not everyone will pick up that they’re talking to a robot, even if the assistant announces that is indeed automated.

“Not everybody is as tech savvy about it, even if you say, ‘I’m a robot,’ ” says Banafa. “How many people will that click with?” Banafa also wonders about the potential for hacking the system — someone with nefarious intent, for example, taking over the technology. “If I can hack into their system, then I can just get rid of that first sentence about identifying itself.”

Huffman says there are a lot of important questions his team is tackling before the system is available to the general public. “We want to have the right conversations happening about how the technology works, so that people can feel comfortable before we go more broadly with this kind of technology.”

Over the next few months, Duplex will be tested by a select group of users and restaurants. Google will introduce three types of tasks: making restaurant reservations, hair salon appointments and inquiries about holiday hours.

Huffman says it’s not built to do much more than that, and if the assistant is asked questions beyond the task given, it will default to a human service person, who will be standing by to take over the call.