“Uhm,” said the female voice. “Can I book a table for tomorrow?” The question came not from a person, but software called Duplex developed by Google to make phone calls. Before the end of the year, some of the company’s users will be able to direct the bot to call restaurants and book tables on their behalf.
In a demonstration last week, Duplex smartly handled questions from a Google employee playing the role of restaurant worker about details such as the size of the party and the name to hold the table under. Then the bot signed off with a cheery “Ok, great, thanks.” Duplex had started the conversation by announcing “I’m Google’s automated booking service so I’ll record the call,” but it was barely distinguishable from a person.
Google announced today that Duplex will be made available on the company’s Pixel smartphones before the end of the year, in New York, Atlanta, Phoenix, and the San Francisco Bay Area. It will be a feature of Google Assistant, the company’s rival to Apple’s Siri; for now, it will only call restaurants without online booking systems, which are already supported by the assistant.
Duplex’s debut makes a small change to Google Assistant’s capabilities. But it marks another moment in the march of artificial intelligence technology into daily life. Investments in AI by Google and its competitors have made it routine for computers to recognize our speech or faces. But even recent AI-powered services with names and voices, such as Apple’s Siri and Amazon’s Alexa, cannot be easily confused with humans. Software that can passably imitate how people talk, and make its own calls, feels…um…different.
Google CEO Sundar Pichai sparked awe but also alarm when he unveiled Duplex in May in a keynote at the company’s annual developer conference. He played two recordings in which the bot did not identify itself when calling apparently unwitting staff to make bookings at a hair salon and restaurant.
A Google spokesperson told WIRED that the company now has a policy to always have the bot disclose its true nature when making calls. Duplex still retains the human-like voice and “ums,” “ahs,” and “umm-hmms” that struck some as spooky, though. Nick Fox, the executive who leads product and design for Google search and the company’s assistant, says those interjections are necessary to make Duplex calls shorter and smoother. “The person on the other end shouldn’t be thinking about how do I adjust my behavior, I should be able to do what I normally do and the system adapts to that,” he says.
The experience of WIRED writer Lauren Goode, who answered a call from Duplex in a demo this past June, illustrates how bots that sound like people can be disorienting. She confused the bot by lobbing a question about allergies in the middle of a discussion about available times for a restaurant reservation. Goode became confused herself when she learned that a second voice that came on the line to complete the derailed transaction was a human call center worker, not another Duplex bot playing clean up.
The term computer was originally applied to people, who carried out calculations manually. Then computers became room-filling machines, then desk-sized, then pocketable. Now they can sound and converse like people, at least in the confines of a dialogue with a very specific goal. “It feels odd because people have this notion that people and machines are different,” says Jeff Bigham, a professor at Carnegie Mellon University who researches human-computer interaction.
Restaurant staff will be the guinea pigs for what happens when that distinction is eroded—at least for certain kinds of phone calls.
Fox, the Google exec leading the project, pitches Duplex as a win-win. Google users will be freed from having to make phone calls to plan their outings; restaurants without online booking systems will gain new customers. “Those businesses lose out because people say ‘Unless I can book this online I’m not going to book,’” he says.
Some people closer to the restaurant business worry that Duplex might make calling restaurants too easy for Google users. Gwyneth Borden, executive director of the Golden Gate Restaurant Association, a trade group for Bay Area restaurants, says people may use the technology to book multiple reservations and then flake out, or call restaurants over and over.
When Borden spoke with WIRED Friday afternoon, her organization had not heard anything from Google during its testing of Duplex or in advance of its coming launch. “If you truly believe this is going to be helpful, why not work with us?” Borden says. A spokesperson for Google said the company plans to begin contacting business organizations.
Restaurants can opt out of receiving Duplex calls by speaking up during a call from Duplex, or through the website where businesses can manage listing information shown in Google’s search and maps services. When calls go awry—Fox says the “overwhelming majority” work out fine—the software will alert an operator in a Google call center who takes over.
Duplex is not Google’s only effort to develop software that talks on the phone. Earlier this year, the company’s cloud division launched tools that help businesses build automated call center software using similar voice synthesis technology to that used in Duplex. Google announced today that its assistant will soon be able to screen calls on Pixel phones. If the feature is turned on, callers will hear an unmistakably synthetic voice asking them to describe why they are calling. A live transcript of what the caller says will appear on the phone’s screen, so the recipient can decide whether to pick up, or call back.
Duplex is notably more ambitious than those other projects. Google plans to iterate fast as it watches what happens when the bot starts making calls in large volumes. One open question is whether the male or female versions of the bot being tested turn out to be more effective. If the initial rollout goes well, hair salons will probably be the next in line to get the Duplex treatment. Google has also experimented with having the bot inquire about holiday hours.
Bigham, the Carnegie Mellon professor, and others watching Google’s project say it probably won’t be the only one with human-like phone bots for much longer. Apple, Amazon, and many smaller companies have launched widely used voice assistants of their own. The impressive voice synthesis technology at work in Duplex is based on research from Google and Alphabet AI labs that has been published openly.
The tens of millions of robocalls placed each day in the US suggest not all the uses for Duplex-style technology would be welcome. Today’s robocalls typically just play a recording; some scammers use human staff. Phone bots capable of back and forth conversation on even a narrow topic could be both cheap and effective. “As this technology gets better it seems completely reasonable that the next guy who phones me trying to convince me to give him my credit card number is not a person or recording, it’s a rogue Duplex-style agent,” says Bigham.
Roman Yampolskiy, director of the cybersecurity laboratory at the University of Louisville, hopes legislation requiring human-like phone bots to identify themselves might moderate how businesses deploy them, pointing to how California recently passed a law requiring bots on social platforms to identify their true nature. He also thinks nefarious uses of such technology are inevitable. “You can use this for sales, you can use this for social engineering attacks,” says Yampolskiy, who recently published a book on AI safety and security. “People will find ways to use this technology that we can never anticipate.”