The killer app

Should we worry that Google's Duplex can talk like us?

Back in 1950, computer pioneer and mathematical genius Alan Turing proposed a simple test of computer intelligence: a computer should be able to carry out a natural language conversation with a human being comfortably enough to convince the other party, and any observers, that it was human. Not surprisingly, despite massive advancements in science and technology, this barrier has not been breached. However, earlier this month, Google demonstrated a new tool — called Google Duplex — that appears to pass the Turing test. At the I/O 2018 developers’ conference, Sundar Pichai demonstrated how Duplex, a virtual assistant (VA), could book appointments with hairdressers, make enquiries about restaurant timings and indeed make reservations using voice-calls made to human beings. The remarkable thing is that the voice-calls, in terms of tonality and speech patterns, are indistinguishable from those made by a human being. It is not easy to achieve this because humans talk in varying accents. They pause, repeat themselves, go off on a tangent, mishear, interrupt, ask for clarifications, and so on. Humans also use tonality to convey meaning and context when a statement may be a joke, a sarcastic comment, or entirely serious. 

Duplex is integrated with Google Assistant and can just be ordered to make voice-calls by speaking to Google Assistant. It is narrow focus and experimental at the moment. It can perform only these “closed domain” tasks of making restaurant reservations and hairdressing appointments and discovering working hours for restaurants. At this instant, it can handle most interactions in those areas, and it has a choice of six voices. It can fill in conversational gaps with lifelike “hmms”. But even a limited skill set could be extremely useful for many businesses, such as an online food delivery service. It also comes with some safeguards: In case it flounders, Duplex can hand over the conversation to a human being. Obviously, Google intends to scale up the number of closed domain tasks that Duplex can perform. It will probably be released in a limited way to a limited number of users this summer. 

Performing these three tasks requires what artificial intelligence (AI) researchers call “real-time supervised training”. This actually means a human supervisor monitoring the VA as it calls, and programming guidance and explanations to the AI as required. This “sitting over the shoulder” continues until the VA can perform autonomously. This is exactly how call centres train new (human) employees and it should send a chill down the spine of the information technology enabled services (ITES) industry. 

Call centres perform similar “closed domain” tasks, such as logging complaints and filling in forms. As Duplex acquires new conversational skills, it could move into these domains and that could trigger a death spiral for a certain kind of call centre. 

The demonstration triggered large doses of outrage and suspicion, along with delight.  Google seems to have redacted parts of the conversations it released, probably to protect the privacy of the establishments that Duplex made calls to for the demo. It is not clear if the system disclosed it was a robot. It does raise ethical issues, including those of privacy violation and informed consent. One can imagine scenarios where a spammer sets up a system of automated calls, or pranksters hoax the police with emergency calls. Human beings will also have to develop social norms to cope with this, especially if known voices can be faked. People might, for example, need to agree on code phrases to reliably identify friends and family from an AI that can imitate friends and family. 

Google agrees that there is an obligation to reveal that a machine is speaking and it claims to have built-in safeguards to prevent some types of spamming, such as repeatedly calling the same number. However, other researchers will surely be looking to build similar tools and there will be a need for sensible regulation to prevent hackers from modifying and misusing such systems with impunity. Once chatty AI systems become common, human discourse could get more complicated.