ALBANY, Calif. — On a recent afternoon at the Lao Thai Kitchen restaurant, the telephone rang and the caller ID read “Google Assistant.” Jimmy Tran, a waiter, answered the phone. The caller was a man with an Irish accent hoping to book a dinner reservation for two on the weekend.
This was no ordinary booking. It came through Google Duplex, a free service that uses artificial intelligence to call restaurants and — mimicking a human voice — speak on our behalf to book a table. The feature, which had a limited release about a year ago, recently became available to a larger number of Android devices and iPhones.
The voice of the Irish man sounded eerily human. When asked whether he was a robot, the caller immediately replied, “No, I’m not a robot,” and laughed.
This is how the call went:
“It sounded very real,” Mr. Tran said in an interview after hanging up the call with Google. “It was perfectly human.”
Google later confirmed, to our disappointment, that the caller had been telling the truth: He was a person working in a call center. The company said that about 25 percent of calls placed through Duplex started with a human, and that about 15 percent of those that began with an automated system had a human intervene at some point.
We tested Duplex for several days, calling more than a dozen restaurants, and our tests showed a heavy reliance on humans. Among our four successful bookings with Duplex, three were done by people. But when calls were actually placed by Google’s artificially intelligent assistant, the bot sounded very much like a real person and was even able to respond to nuanced questions.
In other words, Duplex, which Google first showed off last year as a technological marvel using A.I., is still largely operated by humans. While A.I. services like Google’s are meant to help us, their part-machine, part-human approach could contribute to a mounting problem: the struggle to decipher the real from the fake, from bogus reviews and online disinformation to bots posing as people.
Here are the results of our experiment.
Google’s A.I. is eerily human, when it works
To test Google Duplex, we used a pair of Google’s Pixel smartphones, which include the company’s virtual assistant by default. (Apple’s iPhone users can try Duplex by downloading the free Google Assistant app.) At the bottom of the screen, we pressed a button to summon the Google assistant and then said, “Book me a dinner reservation.”
Google’s assistant then loaded a list of nearby restaurants. For the restaurants that took reservations only over the phone, Google offered to step in and place the call with Duplex.
We tried using Duplex more than a dozen times. Several restaurants, like Henry’s Hunan in San Francisco and China Village in Albany, Calif., rejected our requests for a table of two to four people because they took reservations only for tables of 10 people or more.
Eventually, we secured four reservations in Albany: two separate reservations at Nomad Tibetan restaurant, one booking at Lao Thai Kitchen and one reservation at Bowl’d Korean Rice Bar. We witnessed or reviewed each of the phone calls, and the restaurants were made aware that we were testing Duplex before they picked up the phone.
Only the reservation at Bowl’d, which we witnessed at the restaurant, was made entirely with Google’s A.I. service. The bot introduced itself as Google’s automated booking service, followed by a request to book a table for Tuesday the 21st.
The call demonstrated the ability of Google’s A.I. operator to insert pauses and “ums” to mimic a human — in effect making the interaction feel more lifelike and less scripted. Hear for yourself:
At several moments during the call, the restaurant’s manager, Jin Park, acted confused and asked the caller to state the party size and reservation date. The bot patiently answered the questions again and again. Then, Mr. Park threw a curveball: “Are there any kids?”
The Google bot was quick to improvise: “I’m actually booking on behalf of a client, so I’m not too sure,” he said.
Real Life. Real News. Real Voices
Help us tell more of the stories that matterBecome a founding member
“Everything was perfect,” Mr. Park said in an interview after conversing with the Google bot. “It’s like a real person talking.”
Mr. Park added that he was especially impressed with how the bot handled his question about whether there were children in the party.
But our experience with the other bookings was less impressive as they were all handled by humans. Google said that Duplex was sometimes relying on people in part because it was taking a conservative approach to be respectful toward businesses. Google will have a human involved in the call in a number of situations, like if the company is unsure of whether the business takes reservations, or if the user of the assistant might be a spammer.
Valerie Nygaard, a product manager working on Duplex, said that for our reservation at the Tibetan restaurant, the company might have had a person place the call because it lacked signals indicating the restaurant took reservations.
The next day, however, we tested Duplex at the same Tibetan restaurant, and it again used a human caller despite our earlier, successful booking. So Duplex doesn’t appear to learn quickly.
Duplex needs lots of data to improve
In recent years, the development of A.I. has accelerated thanks to what are called neural networks, complex mathematical systems that can learn tasks by analyzing vast amounts of data. By analyzing thousands of dog photos, for instance, a neural network can learn to recognize a dog.
This technology has significantly improved a machine’s ability to recognize spoken words, understand how these words are used and even generate speech on its own. With Duplex, Google is combining these various tasks into a single system. It works because Google has focused on a small domain: restaurant reservations.
Building this kind of system requires large amounts of data, and Google may be using human callers to generate data that can help “train” future versions of the system.
Nick Fox, the Google executive overseeing its assistant, said the company was not aggressively trying to eliminate human involvement from Duplex, because that could make the experience for business owners worse. Instead, he said, Google was trying to improve the automated system over time and slowly decrease the need for humans to intervene.
Not so smart after all
In an era of tech companies trumpeting the arrival of artificial intelligence, today’s technology is not quite as intelligent as it might seem.
Mark Zuckerberg promises that A.I. can identify and remove toxic content from Facebook, but his company still employs thousands of humans to do the job. When Amazon boasts of all the robots in its distribution centers, blue-collar workers are sorting through all the stuff moving through those giant warehouses.
Duplex is proficient at making a restaurant reservation over the phone, but much like Facebook, Google still leans on human intelligence. At any given moment, it is lifelike. But it struggles to deal with the unexpected.
“There are three things that are important when it comes to A.I.’s interactions with humans: context, context and context,” said Jerry Kaplan, author of “Humans Need Not Apply: A Guide to Wealth and Work in the Age of Artificial Intelligence” and a Stanford University lecturer on artificial intelligence. “Machines are very good with detail but terrible at context,” he said.
The general public doesn’t always see that bigger picture, partly because of the way companies market their technology. When Google first unveiled Duplex last year, it demonstrated making phone calls to a hair salon and a restaurant — handled with artificial intelligence, without a mention of human intervention. The message was all about the power of the technology.
“This is like the footnotes in the TV ad,” Mr. Kaplan said. “The footnotes contradict the marketing message.”
At the same time, the technologies that underpin services like Google Duplex are improving at a remarkable rate. In time, it will become harder and harder to understand what is automated and what is human.
But we need to think long and hard about how this technology should be used and how it shouldn’t. Sorting through all the questions is difficult enough. It gets even harder if we don’t have a clear understanding of what the technology can do.
Subscribe to the newsletter news
We hate SPAM and promise to keep your email address safe