Propaganda reality check: The ‘AI’ phone call and Google Duplex mania

In the wake of a potentially game-changing AI demo at Google Duplex recently, Nico Neumann considers if Google will ever truly be able to provide the perfect hairdresser's appointment via artificial intelligence.

So everyone is going crazy about Google Duplex and the “AI” phone call. I don’t know how many people I have seen sharing the video today with comments along the lines of “this is the future”.

Here are my five questions after watching the clip:

1) Why is Google not showing a live demo? A recorded video could be fake or staged – ironically even using ‘AI‘.

2) Nice sales propaganda, but how many calls didn’t work out until they got these more or less nice examples for the sales pitch?

3) WTF does “gotcha” mean in the second call? I’m very curious what the assistant will report back to you here.

4) We haven’t even nailed chatbots 1.0. I am not convinced this will lead to efficient outcomes in most real-world applications. How often does this save time? What if I get unexpected answers? What if my preferred appointment or option is not available?

Take the following scenario, for example:

You to Google Assistant (GA): schedule hairdresser appointment for THU, 3pm? (not available)

GA to you: appointment booked for THU 10am.

You to GA: rebook hairdresser appointment for Friday, 10am.

GA to you: appointment booked for FRI 11am. (again not available).

You to GA: …

… there are a million scenarios I can imagine where the communication between a human and an automated assistant would take more time than quickly calling yourself.

5) Ethical question – should it be required to reveal that someone is an automated caller and not a human?

Bear in mind what Google writes, (with my highlights):

System operation

The Google Duplex system is capable of carrying out sophisticated conversations and it completes the majority of its tasks fully autonomously, without human involvement. The system has a self-monitoring capability, which allows it to recognise the tasks it cannot complete autonomously (e.g., scheduling an unusually complex appointment). In these cases, it signals to a human operator, who can complete the task.

Please don’t get me wrong. I don’t want to ridicule the impressive progress we have seen in computer science using new algorithmic decision systems. There are some great achievements based on recent machine-learning techniques, in particular for image and speech recognition and various game simulations.

However, nearly all these situations represent cases with a finite, deterministic set of possibilities and outcomes. The rules of Go are clear. Whether an image is a cat or not is clear too.

But what about the perfect hairdresser appointment? This will depend on your personal situation and context. And the conversation can easily go in a direction which an automated assistant cannot handle anymore.

The result will be: lost time, as I may need to call anyway, and frustrated service providers. Think about the shops and places that gets lots of calls without possible bookings or cancellations because of misunderstandings.

In other words, many real-life scenarios include unforeseeable cases that can easily complicate automated processes, as per my example above. I personally haven’t even had a good experience with chatbots yet. Too often you are forced to answer a million questions and in the end you are asked to call the service line anyway.

Getting excited about future opportunities is fine, but we need to think things through and consider reality. Showing one or two examples in a sales pitch is one thing – rolling out large-scale applications in the real world is a totally different beast.

Nico Neumann is assistant professor at Melbourne Business School.


Get the latest media and marketing industry news (and views) direct to your inbox.

Sign up to the free Mumbrella newsletter now.



Sign up to our free daily update to get the latest in media and marketing.