What the Google Duplex debate tells us

The internet had a lot to say when Google demoed its experimental AI voice system, Duplex, at its annual developer conference in Mountain View. In what has become something of a viral video, CEO Sundar Pichai proudly invited us to hear the assistant place calls to make reservations at a hair salon and a restaurant. For a lot of people, the footage heralded a worrying shift in the current AI-human dynamic.

Indeed, many of us were immediately taken aback by the disarmingly human tone and tempo of the software’s artificial speech, peppered with realistic flaws and disfluencies. But once the initial wave of awe and amusement had subsided, and we were done asking if Duplex had passed the Turing Test (the answer is “no”), there was a second wave of social commentary expressing unease about the potential applications and implications of this arrestingly human-like system. It’s hard to understand how Google was naïve enough not to see it coming.

The questions asked were valid ones. Could this tool be used by malicious actors? Wasn’t this a deception? An insult to the dignity of the human on the end of the line? What will remain of our humanity once we’ve delegated all such tasks to AI? Perhaps there was even a realization that we are truly facing the end of human exceptionalism, creating things that can match and potentially exceed our defining capabilities.

Duplex isn’t the only one

There are at least two notable things about the public reaction to Duplex. The first is its sheer force. Articles, bylines, tweets, and blogs articulated and rearticulated the ways in which Google’s latest offering could compromise our valued ethical principles — like honesty and respect. The second is that this avalanche of criticism was so squarely leveled at Google. After all, they are hardly the only bot developer trying to replicate natural conversation.

Take Microsoft. Just a day or so before the Duplex debacle, they quietly launched their Speech Devices software developer kit which allows sophisticated voice assistants to be placed in pretty much any accommodating device. At least in theory, this could signal a near future in which intelligent conversational bots are positioned in our cars, at our work stations, in our showers, in our children’s toys. But this ubiquity isn’t, apparently, what frightens us.

Microsoft’s Xiaoice has more than 500 million users. (Image credit: Microsoft/YouTube)

Just a couple of weeks later in London, we find Microsoft demonstrating another assistance product that feels a lot like a Chinese version of Duplex. Known as Xiaoice, which translates as “little Bing,” this AI already chats with its 500 million “friend” users through messenger app WeChat, and it can now call them to speak on the phone. Despite the high level of humanization going on here, fear levels in response to Xiaoice have been low. So why is it that Duplex unleashed the full force of our bot rage if this similar software is considered fun and inoffensive?

Well, it could be that the Chinese are typically less daunted by AI. Or that this specific AI is simply less daunting, gently giving its users instructions to “close the window before you go to sleep,” and the like. Or perhaps it’s because English-speaking audiences aren’t very good at assimilating the potential of non-English speaking conversational products.

The muted response of our mainstream media likely resulted from a combination of all of these factors. But I also believe the answer is bound-up with the concepts of knowledge and authenticity.

Knowledge and authenticity

If this sounds a little abstract, that probably won’t be the case for too long. In an era of “fake news” that deliberately misinforms us, bots that are indistinguishable from humans, and the creation of deep fake videos that cause us to distrust our most basic “seeing is believing” instincts, determining the truth is becoming paramount. At a recent tech trends event in San Francisco, venture capitalist Tomasz Tunguz speculated that the “hunt for authenticity” would become a growth area in the next three to five years, delivering us a new class of software.

In this dichotomy, it’s not hard to see that Google Duplex is part of the problem rather than the solution.

But isn’t Alexa similarly “inauthentic”? Aren’t Xiaoice’s conversations as phony as anything Google has produced? Well yes — by definition — but this is where the importance of knowledge plays in. Neither Alexa or Xiaoice hide that they are an AI. When I solicit their advice or companionship, I do so in the full knowledge that they aren’t human. I do not require that they prove their authenticity. In contrast, when Duplex proactively calls me up to book theater tickets at the box office I work for, there is a clear information imbalance — in short, Duplex makes me a dupe.

Amazon’s Alexa voice assistant aims to integrate itself into daily living. (Image credit: Amazon)

As we march further into a world in which both the human-AI and the online-offline distinctions are blurred, we’re going to have to ask whether we are comfortable chasing this kind of dupe in its many guises. If whatever convenience it yields is worth really the price. Just how important is it that our conversational bots sound exactly like real humans? Particularly in scenarios where they make an approach?

Google’s response to the visceral public reaction has been contrite. If-and-when Duplex moves on from its current experimental phase, it will disclose the truth to the real humans it calls. This move could even defeat Google’s own purpose in creating the software, which is to “enable people to have a natural conversation with computers, as they would with each other.” Or at least it could if this admission leads us to deal with AI conversation in a different manner to our human interactions. Whatever the case, the episode has demonstrated how highly we value this transparency, and developers should be paying close attention.

Domain specific

Lastly, although we clearly need to choose the right moral path here, there is an important caveat that concerns the highly limited use of Google Duplex. Though in theory the scope for misuse and deception is troubling, in reality the product Pichai showcased would struggle severely outside of the “closed domains” within which it currently operates — reservations at restaurants or hair salons. As yet, there is no need to become paranoid about bots acquiring artificial general intelligence and ensuing dystopias. It would be remiss not to note this.

Furthermore, the domains in which conversational bots can be useful are, in themselves, open to dispute. Speaking only for myself, it’s been a long time since I have called either a salon or a restaurant to make an appointment. These tasks were replaced long ago by button clicks and easy-to-use apps. So we shouldn’t just assume that the future is one in which minimum wage workers are engaged by the virtual assistants of the 1%. It seems likely that hair salons and restaurants will also have assistants. These assistants will interact with our assistants, performing administrative tasks in business environments where the “dupe” matters less.

In the context of recent conversational AI developments, we should still heed the fact that Duplex, arguably a most outwardly impressive feat, has caused this particular ethical debate. Bot developers could do worse than examining the trigger factors, and considering how to bake transparency into their conversational products.