èƵ

Google’s human-like phone calls are a clever but nasty trick

Google’s AI assistant is getting an upgrade to let it make you an appointment by impersonating a human on the phone, but this tech could cause a host of problems
A woman answers the phone
May I ask who is calling?
MBI / Alamy Stock Photo

Get ready to spend your life talking to a robot. Yesterday, Google’s CEO Sundar Pichai demonstrated a technology called , which can make phone calls on behalf of users – and sounds like a real human.

He played two examples in which the system made an appointment for a haircut and phoned a restaurant to make a reservation. Unlike the robotic phone calls we are all used to, Duplex does an uncanny impression of human speech, complete with ums and ahs and well-timed pauses in the conversation. It also appears to understand what the person it is speaking to is saying, and responds deftly when the conversation doesn’t go as expected.

It is a very impressive demonstration. But clearly Google is showing us the best examples of its calls; we don’t get to hear the missteps. And the system is still very limited: the neural network underlying it, which was trained on real phone calls, has only learned to master conversations to make these kinds of appointments – it won’t call your mother for a catch-up chat. If you have ever interacted with a voice assistant like the Amazon Echo, you will know that language processing technology can do some things very well, but struggles to deal with the huge variety of things we say to each other in natural conversations.

What happens when this technology takes off? Duplex might save us the minor inconvenience of booking a restaurant, but when technologies like it fall into the hand of marketers and scammers, the outcome could be a deluge of spam calls that can’t easily be identified as such.

Caller unknown

By teaching the system to sound natural, with all the quirks of human speech, Google says its aim is “to make the conversation experience comfortable”. But those impressive ums and ahs are designed to deceive – if the AI falters and the listener realises they have been hoodwinked, it will backfire badly, as no one likes to be fooled. A transparent approach would ultimately be more fruitful: if we are told upfront that we are talking to a robot and the conversation runs smoothly, such interactions would soon feel normal.

The next concern is that Google is recording a large number of phone calls to train the technology,  but it isn’t clear how the data has been gathered or whether the people being called have given their consent. The examples played in the demonstration didn’t include any indication to the recipients that they were being recorded. Google says it is working on this. “We want to be clear about the intent of the call so businesses understand the context,” a . “We’ll be experimenting with the right approach over the coming months.”

It wouldn’t be the first time that Google’s data-gathering practices were found to be inappropriate. There might be little harm in this when the phone calls are to book a haircut, but it would be another matter if Google were to expand Duplex’s abilities to book medical appointments, for example, and sensitive information was being disclosed.

Google plans to begin testing the feature in its Assistant app in the summer. Before then, it needs to be much clearer to those on the receiving end about the nature of the caller and how recordings will be stored and used. Otherwise, the next time the phone rings, you might not know if you are talking to a human.

Topics: Google / Machine learning