Your browser is not supported. Please update it.

23 May 2018

Google Duplex is an AI assistant which bridges digital and physical worlds

It seems counterintuitive, but Google’s new and somewhat uncanny Duplex system is all about helping consumers use Google less. The new assistant, which was demonstrated making calls to book a haircut, is designed to connect the Google and Android digital world with the analog world – calling fleshy humans on its user’s behalf, for that frictionless experience.

The end goal here is to create better customer experiences inside Android and the wider Google ecosystem. It’s an extension of Assistant, intended to work behind the scenes once a user asks ‘Hey Google, book me a table at a nice restaurant tonight’. In theory, this would be seamless, and involve a single voice command – rather than the manual process of searches, calls, and then calendar entries.

However, there has been a lot of humming and hawing about the ethics of phoning up a real person. But this would not be a concern if such businesses had some form of reservation API, or a form that could be filled out on a website. We’ve seen similar features from fast food vendors, with Domino’s going as far as creating those Amazon Dash buttons for rapid ordering of pizza – so we don’t place all that much weight in the apparent upset at Duplex’s function.

It would be better if Duplex had identified itself in the demo as a digital assistant, so that the person on the other end would be in the know. As the conversations are recorded, there are some concerns that not informing the recipient would violate telephony and privacy laws in some jurisdictions, so Google should be wary of this.

Google is also going to some length to obfuscate Duplex – adding in ‘umms’ and ‘ahhs’ to make it sound more lifelike. To this end, it nailed the Turing Test, as the person on the other line had no idea it was not a real person, and while it is very impressive, there will always be people with reservations about this reservation system.

In Google’s view, it is a means to an end, as CEO Sundar Pichai said that 60% of businesses in the US “don’t have an online booking system set up. We think AI can help with this problem,” before also mentioning that it would be used to check opening times that can be confusing during holiday periods. Google would aggregate such calls, crowd-sourcing that knowledge to share through its platform – so these businesses might actually receive a very low volume of such calls.

It might also be the kick that many businesses need to modernize their online presence, and while Duplex is starting out in the service industry, it wouldn’t be a stretch to perform functions like stock checking and reservation – although retail has generally been better at making those features available online. Like restaurants, however, there is a huge volume of smaller traders that have not yet done so.

Google will have to ensure it takes sufficient steps to avoid Duplex being abused. Robocallers would love to get their hands on something this sophisticated, but even at an end-device level, drunken commands or pranks from friends could have very real consequences for small business owners – fooled by an apparently earnest AI system in to an action that might hit them in the wallet.

In terms of IoT ramifications, this is a way of bridging the real world with the digital one, conveyed via a user’s smartphone. Expanding upon this, Duplex might only be a transient flash in the pan, until those businesses catch up and create their own web forms or APIs, or get amalgamated into some sort of platform that could be polled.

Should Google widen Duplex’s remit, which we expect it will, then this should give users a way to interrogate any such online system. The smartphones are a unifying force to some extent, and should lead to developers consolidating around a pretty standard approach to interfacing with things like Duplex. This should make IoT sensors just as reachable as your local Chinese restaurant, all via a one-line voice command to your smartphone – but that future is still a long way out. Such consolidation takes a long time.

In related news, Google launched ML Kit – an SDK to let developers perform machine-learning functions on end-devices rather than in central clouds. It also unveiled a few details about its new TPU 3.0, the accelerator chips that Google designed for use in TensorFlow operations.

Pichai, publicly at least, is taking a cautious approach for AI and ML at Google. He added “technology can be a positive force, but we can’t be wide-eyed about the innovations that technology creates. Very real questions are being raised about the impact of advances and the role they play — the path forward has to be calibrated carefully.”