Last year, when Google first introduced its new Duplex feature, which used artificial intelligence to make calls to users, it really scared us because it sounded so human.
During the demo, CEO Sundar Pichai showcased Duplex as a fully automated service. Google’s AI assistant Duplex called a hairdresser to schedule an appointment and called a restaurant asking for information. The automated voices behind the calls were remarkably lifelike, with the voices also featuring some very human-like vocalizations.
The calls were so human-like that it seemed almost impossible for any listener on the other end of the line to determine whether they were human or AI.
A year after Duplex launched, it's even harder to make that distinction – because Duplex calls could actually come from either a human or an AI.
Google confirmed to the New York Times that about 25 percent of calls made by Duplex start with a human, and about 15 percent of calls initiated by AI get human intervention at some point during the call. That means just 60 percent of calls made by Duplex are completely automated.
However, the Times conducted its own investigation, using Duplex to call more than a dozen restaurants, and the Times found that three out of four successful Duplex reservations were made by humans.
Nick Fox, Google Assistant’s vice president of design, told the Times that Google was slow to eliminate human involvement with Duplex because the company didn’t want to cause a problem for businesses that need to answer calls. Fox explained that Google wants to progressively improve the feature and reduce human involvement year after year.
As the Times points out, building a system that is completely automated requires a lot of data to allow the neural networks to be ready for any difficult or confusing questions during calls.
Google is likely using humans on calls to create and aggregate this data to teach Duplex.
