Google prepares to roll out its phone-calling AI in the United States
Silicon Valley's quest for artificial intelligence has led it to build self-driving cars, drones, and robots that can do backflips. But often that journey has come down to something much more prosaic, such as ordering a pizza, or booking a restaurant reservation.
Source: Elizabeth Dwoskin
That's how I found myself sitting around a telephone at Oren's Hummus Shop, a cafe near Google's campus in Silicon Valley, on Tuesday afternoon; the setting the company had chosen for the first public testing of Duplex, its new artificial-intelligence-powered conversational assistant.
Duplex is the company's next-generation virtual helper. When the company first showcased it at its developer conference in May, it engaged in conversation so lifelike — complete with humanlike "ums" and pauses — that the person on the other end of the call couldn't tell that the speaker was just software. Some asked whether the interaction was fake. (Google said it wasn't.) Others raised ethical questions about whether Google's assistant should disclose in conversation that it's a machine, and that it may be recording the call.
In the months ahead, Duplex will be introduced for three types of tasks: booking restaurant reservations, calling stores to inquire about holiday hours, and making hair salon appointments. Ask it to make a dentist appointment or anything else, and it will be baffled. A small group of beta testers will get to use the assistant, along with pre-selected businesses, across the United States.
At Oren's, a group of reporters took turns using the telephone, pretending we were Oren's employees tasked with taking people's reservation requests.
"Hi! I'm calling to make a reservation," a very realistic-sounding woman's voice said when I picked up the phone. "I'm Google's automated booking service, so I'll record the call."
"What's that?" I asked.
The software couldn't hear me at first — even a little background noise can throw off AI — and it barrelled into a reservation request. I repeated the question more loudly.
"I'm making a reservation, for a client, for Monday the 2nd," the voice said, unflustered.
The reporters did their best to trip Duplex up, but the technology remained polite and unflappable, and not annoyingly cheerful. Duplex responded to every twist and turn in the conversation, offering to call back later if no spots were available.
Google said the "ums" and "uhs" — parts of speech called disfluencies — were built into the software after the company realised that people were more likely to hang up without them. "We began iterating to make it more natural," said Scott Huffman, vice president of engineering for Google Assistant. "We had a higher success rate." He said the technology can now successfully book a reservation in 4 out of every 5 calls it makes. (When it fails it can cry for help, reverting to a real person working in a Google call centre.)
I asked Huffman, who used to run Google Search, why he didn't go the opposite route and make the voice sound purposefully robotic so that people would slow down and become more deliberate, as if they were talking to a computer. Wouldn't that be less potentially deceptive and yield clearer results? Huffman explained that unlike Google Search, which is a destination people seek out to retrieve information, the goal of Assistant was to fit into people's natural workflow and be the least disruptive as possible.
Duplex has built in new prompts to be more transparent about who's on the line. If asked, "Are you a robot?" the system responds, "I'm an automated system built by Google."
Huffman and a colleague, Nick Fox, said they decided that Google Assistant would disclose this type of information upfront after experts raised concerns that people should have the right to consent before talking to a bot. The executives said that legal considerations were not on their minds when they made the decisions. Two bills - one in Congress and one in California - would require such disclosure.
Duplex is not the first AI to order food effectively through chat or to be so lifelike that it can confuse people into thinking it's human. Viv, an assistant built by Siri's founders, was a nimble pizza orderer when I reviewed it in 2016. And Amy, a virtual assistant I once used to schedule my meetings over email, was such a natural conversationalist that a human secretary who had interacted with her once asked me to wish her happy holidays at the end of 2015. (I explained, awkwardly, that Amy was a robot, and felt embarrassed not to know the right etiquette for this uncharted human-computer interaction.)
But Google's virtual assistant has a lead over competitors because of the vast data collection it uses to train its AI. And that's pushing it into the mainstream faster than anyone else.
"We could sound robotic and expect to be treated like a robot," Huffman said. "Or on the other hand, we could disclose that we were a robot and fit into people's workflows. We thought the latter was better." Over time, he said, "we'll see if we made the right call."