Update: Google announced that they will identify themselves at the beginning of each call Google Duplex does. This doesn't change the point of the article, which is that the company somehow considered the first version of Duplex, which is in spirit a hack, worth releasing to its userbase. As an aside, I really would like Duplex to open each call by the rejected title of this article: "Hello, Yes, This is Goog"
Flowing from the cornucopia of AI tricks that Google presented yesterday, one still managed to stand out: Duplex, a technology that enables Google Assistant to make reservations over the phone. You really should watch the demo (1:55:14 to 2:00:50):
The entire sequence was spectacular. But, without really knowing why, I also found it kind of odd. Nonwithstanding the real ethical issues associated with Duplex (more on that in a moment), it stroke me as being, for a Google product, out of the ordinary. And here's why: Duplex is a consumer product made out of a hack. For a company as established as Google, that is not OK.
I mean "hack" here in an informal sense, as "a piece of computer code providing a quick or inelegant solution to a particular problem." Duplex fits that definition perfectly. The particular problem here, as Google mentioned in their press release, is that most small business take reservations only by phone, and we often are too lazy to call. Google could patiently push small business to use online bookings, but that's apparently too slow-moving. Instead, Google decided to provide automated reservations by leveraging technologies, such as voice synthesis and language processing, that are good enough and readily available (the quick part), even though the undertaking relies on, at least, some mild deceit (the inelegant part).
The issue is that hacks generally don't scale very well, and Google Duplex, if kept in its current form, probably won't be an exception.
First off, what was shown yesterday fundamentally relies on subterfuge, and I am not sure that Google can legally keep doing that for very long. Even though Google likely isn't required to disclose to the businesses that this is an automated system calling, I wouldn't be surprised if there is a push for such mandatory disclosures down the road. More immediately, it is also almost certain that if Google wants to record calls to train their models, it will have to disclose it in certain U.S. states (such as California), and many other countries as well.
Maybe Duplex legally won't have to open all of its conversations by, say, an artificiality disclosure. Still, business owners will surely recognize at least some of Duplex's calls. These will probably be repetitive to an extent. Google has only synthetized six different voices, and they will probably have some recognizable characteristics, whether in vocabulary, tone, or intonations. For instance, in one example, Google Assistant claims it is making a reservation for "a client," which is a sentence small business probably almost never hear. If, as Google hopes, Duplex becomes popular, the nail salons and bowling alleys of this world will be called by the same "people" very often.
So, what happens when small businesses become aware that the person calling to book a haircut is, well, not a person? That is the biggest unknown of Duplex, but Google shouldn't necessarily be optimistic. If I were a maitre d' receiving 20 booking requests a day from a natural-sounding, yet still automated, voice, I would have two reactions:
- The clients using Google Assistant are fucking lazy. If my restaurant is popular, no Google reservations, because you can at least afford to make a damn phone call.
- If part of my day really consists in talking to robots, why should I talk to them normally, and say "thank you" or "hello"? I would just talk to that fake person like I talk to my phone: with queries that demand minimal effort. And maybe I would sprinkle these queries with some abuse and insults, since I could get away with it.
Iâm hyperventilating on purpose here: most businesses will probably be fine with Duplex, the most annoyed ones will have an incentive to switch to online booking, and maybe a few high-end holdouts will refuse it altogether; and all will be fine. But I'm stuck on Sundar Pichai's pride that Duplex had been engineered to pretend it's a human, notably by peperring its sentences with strategically placed "hmmm" and "uuuuh". I mean, Google is a careful company, and I am sure the keynote was vetted by lawyers and PR people. That they could not see the glaring problems of Duplex, both in its nature and the way it was presented, is astonishing. I mean, it should have raised off alarm bells that engineers were getting paid to make human voices sound more natural so that the interlocutor wouldn't guess the caller is actually a computer.
It's only a matter of time before people realize that these calls are automated, and Google should act accordingly by systematically disclosing it. Paradoxically, that yesterday's demo was so smooth showed that Google could, and maybe should have, introduced Duplex a couple of years ago. At the time, the language processing probably was close to ready for prime time. Yes, the voice synthesis wouldn't have been as good, and people would have been able to spot that it wasn't a real person. But that's the way it should be, and especially for a company like Google that has a vested interest in not sabotaging its friendly image.
Yesterday, however, all of these concerns were passed over. The company seems unmindful of the nature of Duplex, even though it is very different from the stuff they usually introduce. Again, Duplex is not a neat new app, but a hack, a technology that is both exciting and impure. Even though it might be an eventual success, its roll-out will be much messier than most of Google products. Honestly, if the company keep being so oblivious the complications of introducting a technology to the real world, and so casual about its potential creepiness, they run the risk of galloping straight towards a wall over and over again.