Google Duplex, the AI Assistant we should avoid using
By: +David Herron; Date: May 9, 2018
The demo's are innocent enough. A human-sounding voice places a call to a restaurant or hair salon, and engages in a conversation pretending to be an assistant and to be scheduling an appointment or reservation. The demos include several roadblocks, such as a restaurant that doesn't make reservations for groups less than five people, which the Google Duplex assistant successfully recognizes and handles. The synthesized voice includes lots of human mannerisms and it's amazing to observe in action.
Is artificial digital assistant technology actually needed?
Here is the demo at Google I/O
The conversations are amazing in that computer software is able to navigate the foibles of human conversation. The Duplex assistant even navigated conversations that went awry, such as a clearly Chinese restaurant worker who barely knew English and whose conversation went sideways. The Duplex assistant handled that call getting to a successful conclusion.
But - the supposed drudgery of making phone calls to arrange an appointment is a FIRST WORLD PROBLEM. Those of us living in the hyper-developed countries perceive this as a problem. It's not actually a problem. Making a call to make an appointment is an opportunity for human contact. Why would we want to avoid making contact with a fellow human being?
The Turing Test
In the early days of computing, it was foreseen the future computer scientists would reach this stage - machines that can mimic humans well enough to fool other humans. One early computer scientist, Alan Turing, proposed a test for the success of artificial intelligence software. A Turing Test involves a human sitting at a teletype conversing with software and getting fooled into thinking they're conversing with a real human.
Methinks Alan Turing was unable to conceive that we'd be able to carry a supercomputer in our pocket, and use it solely to make phone calls, take pictures, post goofy pictures on snapchat, or play games. In any case, these Google Duplex phone calls represent a kind of Turing Test and demonstrates that AI technology has progressed to where the Turing Test has already been met or will soon be met.
This is a kind of turning point, isn't it? AI software that passes the Turing Test?
We all get agitated with stupid voice-driven answering systems, right? Many of us respond by pressing zero ('0') to get to a live human rather than to continue with the garbage interaction with a voice-driven menu system. With artificial intelligence such a system could be more useful.
A busy restaurant might want all hands engaged in dealing with customers, and therefore might want a digital answering robot to take reservations.
A door security sentry at a business might have a database of employee pictures and habits, and do a better job of screening for access than the current system of RFID chips in employee badges.
But let's not get lost in the positive benefits from artificial intelligence digital assistants and ignore the dangers until they bite us in the ass.
Google is claiming they're designing in safeguards, such as disclosure when Duplex places a phone call that it is a synthesized faux-human. That's nice and ethical of Google, but not all organizations are Google. And in many ways Google isn't Google either.
Surely other organizations will shortly be able to implement similar products, and some will not bother implementing the safeguardds. What could go wrong?
Political dirty tricks with robot impersonation
This is the age of Fake News and roboticized fake social media accounts, fake news blocks, fake this, fake that, and we're in the middle of a War on Truth itself. Some countries or companies see it being to their benefit to make us all doubt the truth, and are therefore sowing disinformation in a big way.
What if a digitally synthesized human was created to imitate a given person. The synthesized person could be made to engage in any conversation, even ones counter to what the live human would do. To take an example, Sen. Elizabeth Warren has a powerful enough message that opposing forces ARE running political dirty tricks against her. What if they synthesized an imitation of Sen. Warren to place a call with a corporate representative, to arrange a fake payoff for some political favor? Then the audio of that completely fake phone call were blasted all over Fox News and other outlets, causing Sen. Warren to have to resign in humiliation even though she never did such a thing?
I imagine it would be immensely difficult to defend against such a scenario. Similar tricks are already taken - using scammy video editing techniques, for example, and have been used to bring down some people or organizations. With a roboticized phone call it could be even easier to pull off such a trick.
Robotic political games of other kinds
There's plenty of other possible dirty tricks to play. For example Robocalls meant to convince voters one way or another have existed for years, but rely on a simple recorded script. A roboticized digital thing that can pretend to be human could place the call, have a nice interchange, and be programmed with the exactly correct phrases to say to sway humans.
Massive scale con jobs with robotic sales pitches
Similarly it doesn't have to be political dirty tricks. All kinds of companies are trying to get our attention, and might want the scale of roboticized agents to place those phone calls. An AI robot designed to sell car insurance?