AI-driven Google Duplex still relies on humans manning the phones

Call center
Image credit: FrameStockFootages/Shutterstock

Google Duplex – the AI-powered service that can call restaurants to place a booking on your behalf – remains heavily reliant on human beings.

The free service from Google is available to iPhone and Android users, and the idea – if not the technology – is a simple one. Rather than calling a restaurant yourself to book a table, Google Duplex will make the call for you, using artificial intelligence to hold a conversation with the person in charge of reservations. Except in a quarter of cases, it's a human rather than AI that is placing the call.

The finding comes courtesy of the New York Times which discovered that in a surprising high number of cases, the realistic-sounding human on the other end of the phone really was human rather than the Google assistant.

Google Duplex has been available for around a year now, starting off as a limited release for some Android users before expanding to a wider range of handsets including iPhones.

To many, the idea of a robo-calling system being used to place bookings seemed a little too good (or too futuristic) be true, and for a percentage of calls, this is the case.

The New York Times' investigation into Google Duplex led to a confirmation from Google that for around 25 percent of Google Duplex calls, it was a human who initiated them rather than an AI system. On top of this, a further 15 percent of calls which were started by the AI system required intervention from a human.

Too good to be true?

While these are hardly damning statistics for Google, it shows that Duplex still has some way to go before it can be considered self-sustaining. When Google announced Duplex last year, it was described as "an AI system for accomplishing real-world tasks over the phone" and Google conceded that the system was to be "constrain[ed] to closed domains, which are narrow enough to explore extensively". It added: "Duplex can only carry out natural conversations after being deeply trained in such domains. It cannot carry out general conversations."

This is precisely why the system is only being used, at the moment, for restaurant reservations. There are, in theory, a limited number of queries and responses that could arise in a booking scenario, making it easier for an automated system to deal with. But the findings of the New York Times show that even in this relatively limited environment, there is some way to go.

Google plans to expand the capabilities of Duplex into other areas. It has talked about making reservations at hairdressers, and at its Google I/O developer conference earlier this month, the company talked about expanding it from smartphones and onto the web. CEO Sundar Pichai has suggested using the system to make a car rental reservation, or booking movie tickets, as well as helping people to fill in online forms.

Work in progress

The revelation that so many calls currently require human interaction could be seen as an indication that Google Duplex is failing, but it is more complicated than that. As an AI-driven system, Duplex is learning, evolving and improving all the time, so its success rate should naturally start to improve.

But there is also another side to consider: the human receiving the call from Duplex. While very human-sounding, Duplex-initiated calls are identified as such, and those answering the call will learn over time how best to interact with the system to improve efficiency and success rates.

Sofia Elizabella Wyciślik-Wilson
Freelance writer

Sofia is a tech journalist who's been writing about software, hardware and the web for nearly 25 years – but still looks as youthful as ever! After years writing for magazines, her life moved online and remains fueled by technology, music and nature.


Having written for websites and magazines since 2000, producing a wide range of reviews, guides, tutorials, brochures, newsletters and more, she continues to write for diverse audiences, from computing newbies to advanced users and business clients. Always willing to try something new, she loves sharing new discoveries with others.


Sofia lives and breathes Windows, Android, iOS, macOS and just about anything with a power button, but her particular areas of interest include security, tweaking and privacy. Her other loves include walking, music, her two Malamutes and, of course, her wife and daughter.


You can find her on Twitter and Mastodon.