The Ethics of Google Duplex

Image for post
Image for post

No, it doesn’t have to do with robots calling people and people not knowing they’re speaking to a robot. It has to do with training.

Earlier in my career, I did a lot of lead generation work. I’ve made over 20,000 cold calls with different startups and larger companies, including IBM. For IBM, I was selling application servers and middleware to small and medium size enterprises in southern Ohio and northern Kentucky. Think flower shops to credit unions. Being an inside rep, I’d work with my outside rep to setup sales calls.

One of the first appointments I had scheduled for Ken, my outside rep, was to a pig farm north of Lexington, KY. The technology director showed up for the meeting in mud-covered hip waders. They did not have a need for WebSphere Commerce Server.

Image for post
Image for post
Where I sent my first IBM rep.

At IBM, I was relentless. If there was evidence of me having OCD, it was that I needed to keep score of everything. I had created a game where I needed to get to 100 points per day and I would not be able to leave until I got those 100 points.

Here’s how the game broke down:

A deal closes = 100 points

Submit a bid = 50 points

Send a quote = 20 points

Run a customer webinar = 10 points

Create a detailed response to customer =5 points

Conversation with prospect/ customer = 3 points

Call attempt or email = 1 point

The idea was to align my reward (which was really just relief of my quasi OCD) with a behaviour that would create value for IBM. For the last six months of my work there, I’d be playing to hit my daily 100.

I was well trained coming into IBM for this type of game. At my previous employer, a VC-backed startup out of Ottawa, I dubbed myself the Cold Calling Machine — CCM for short.

All I’d do is call companies that had field staff and setup appointments for outside reps to sell electronic forms as a cellular plan add-on. The incentives for showing traction were high and I’d make hundreds of dials a day.

Over the course of the day, I’d take my first break when I got to 75 dials, eat lunch at 150, another snack at 200, etc. Or, I’d take a break after I hit my first appointment. On average, it’d take about 70 dials to get one appointment and all of it was tracked in an Excel spreadsheet because this could be updated very quickly while on the phone (sub-10 s per update vs 20–30 seconds in a CRM).

My call tracking spreadsheet — CCM. Showcasing my Excel skills in 2006.

Whether it was IBM or other companies, the first few hundred calls were pretty bad.

I tripped over my words, I wouldn’t quite understand the product, I wouldn’t know how the culture of a given industry handles phone calls. Also, when should I cut bait and schedule a call back vs reeling in the appointment? — At the start in every company I worked at, I was pretty lousy to start and would get a bit queasy and sweaty palms.

Knowing that I’d be pretty bad on my first few calls, I tried starting on lists that I knew would be particularly bad. The idea was to preserve the good lists (think the Glengarry leads) for when I’d be better.

Image for post
Image for post
The coveted Glengarry leads.

In hindsight, I realize that there were a lot of people who received an unpleasant call from me. Not because I wasn’t offering something of value to them, which I was, but because I couldn’t communicate it to them effectively.

They had to listen to someone nervous and bumbling try to figure out how their industry worked, their phone culture, and what was of value to them. They didn’t know that the person on the other end of the line was just learning. They had their schedules to run and my call was just an interruption. My redeeming quality, however, was that I was a human.

Image for post
Image for post
Human: It’s the part I was born to play, baby!

This is where we come back to Google Duplex. Machine learning includes learning and you learn from trying a bunch of things, messing up in many, and doubling down on the things that work. When you encounter something new, you try different approaches. It doesn’t matter if it’s a human or a machine, the only difference is in the scale and speed.

The recorded demos of Google Duplex at Google I/O were great. They didn’t frustrate the business, they provided a service to these businesses, they were polite and cordial. They sounded like a young personal assistant calling, which is disarming to the recipient.

However, knowing how many calls it takes for a human to get better, my biggest question is: how many calls did it take to get Google Duplex to have a successful call? What happened on those calls? Was anyone really upset? Did it tie up a businesses line?

Which businesses were Google’s training list (non-Glengarry leads)? Which were the ones that were preserved? Who chose this list?

That’s the real ethical question. Is it acceptable to be training fodder for another company’s AI offering without consent? I don’t know what Google did to train Duplex but transparency around this could earn them either a lot of heat or potentially a lot of goodwill.

Written by

Independent daily thoughts on all things future, voice technologies and AI. More at

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store