Google Duplex is Using Humans to Make Restaurant Reservations and Not Just AI
The New York Times reported yesterday that Google Duplex is using humans to make restaurant reservations, at least in some circumstances. Google Duplex was introduced into four locations in November 2018 and then expanded to 43 states in March 2019. The service can make phone calls on behalf of a user to perform tasks such as making restaurant reservations, setting hair salon appointments, and inquiring about hours of operation for retailers. Making a restaurant reservation is the first of these use cases to roll out broadly and previous reporting confirmed the use of the Google Duplex assistant, an AI-enabled bot, to book tables.
Google Didn’t Reveal the Use of Humans for Duplex
However, there was never an indication by Google that humans were in the loop for any of these activities. Google stunned many observers when it demonstrated the service in May 2018 to its I/O developer conference audience. The humanlike sound and speech patterns were better than people expected and the AI appeared to be agile enough to manage even non-linear conversation as long as the domain, such as appointment setting, was limited. Google’s goal is to take care of mundane tasks for users so, in theory, a user would not care whether it is a human or an AI-based bot executing the activity. Using a human is an expensive way to complete the task. The bigger story is that Google was not forthcoming about the state of the technology.
In some ways, it is surprising that it took this long for the practice to be exposed. The reason for this is likely that the speech and conversational capabilities of Google Duplex are so strong that people may simply not have noticed when humans were inserted into the process. You can hear from the recording published by the New York Times that the caller to Lao Thai Kitchen in Albany, California has a noticeable Irish accent and the restaurant host even makes mention of it. Although not in the recording the article reported that “When asked whether he was a robot, the caller immediately replied, ‘No, I’m not a robot,’ and laughed.” You can listen to a portion of the call below.
Filling in Gaps and Amassing Training Data
The New York Times reporters confirmed with Google that the call was placed by a human. “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.” That means about 64% of all Google Duplex tasks are fully executed by the AI and 36% require some sort of human assistance. However, the Times reporters ad hoc testing with over a dozen restaurants resulted in four completed bookings with three made by humans. Those results suggest something closer to the inverse with 25% of calls using Google Duplex. According to the Times:
In other words, Duplex, which Google first showed off last year as a technological marvel using A.I., is still largely operated by humans.
Google told the reporters that it was taking a conservative approach because the system doesn’t yet know which restaurants accept reservations. Much of this information is actually accessible in the Google Knowledge Graph and through integrations with third parties so that doesn’t seem like a great excuse.
It is, however, interesting that the script for the human was similar to what you hear when the Duplex bot calls saying first it is from Google and that the call may be recorded. The warning about the recording is required in several states, including California where this call took place, where all parties on a call must consent to the recording. A key difference is that the human actually moves through the script a bit faster than Google Duplex. These use of a similar script and humans could also become training data for Duplex to use in its efforts to improve over time. Both humans and bots face similar challenges when conversations do not proceed as expected so anticipating those differences based on real-life human-t0-human calls could accelerate the effectiveness of Google Duplex.