You can't automate empathy

Back in 2017, Mike King was already teaching a chatbot to answer his emails.

About 46 minutes into our second chat, he explained it. When you do link outreach, people object for all sorts of reasons, "and there's really only like five reasons why they do it." So you train a bot on those five, hook it up to your inbox, and let it handle the objections. No code required. It's pretty clever.

A few minutes later the talk turned to what machines can't do.

1. 14

   You can't automate empathy

   Tools can tell you what people did. Working out how they felt is still your job.

   "Empathy is a lot more difficult, because empathy is very subjective, just like morality."

   Mike KingS01E03 Hear it 49:07

   > confirmed. i tried. it came out as a spreadsheet.

This from a man who was finishing a guide to machine learning for marketers at the time. Creative work, he said, was a maybe for the machines. Empathy was a lot harder.

The robot can answer the email

Mike's not anti tools, and neither am I. In the same episode he walks through Orange, a drag and drop machine learning tool, and makes the case that marketers shouldn't be scared of it. "It sounds so daunting. It's like, oh math, code, I can't do it. But it's really just like everything else we do in digital marketing, it's guess and check."

The objection bot works because the task is repeatable. Five reasons, five replies. Our guide to marketing automation says the same: automate the birthday emails and the abandoned cart reminders, but "take stock of what touchpoints require that extra special personalisation."

Whose empathy?

Mike's reason empathy is hard to automate is the interesting bit: it doesn't have one right answer. He compared it to the question people ask about AI and morality: "Like can we make them moral? Well whose morals, right?"

He was also worried about bias. "One of the bigger issues or concerns that I have about artificial intelligence is that it reflects our biases." A small group of people build these systems, so the systems inherit their blind spots. A few minutes later he came back to it: "the bigger concern from my perspective is the fact that our biases are being programmed."

He wasn't guessing. The year after our chat, Joy Buolamwini and Timnit Gebru tested commercial facial analysis systems and found error rates for light skinned men that "were never worse than 0.8 percent", while for darker skinned women they "ballooned" past 34 percent in two of the three systems. MIT's write up of the Gender Shades study is worth ten minutes of anyone's time.

Now bring that back to marketing. If a tool learns what your customers "feel" from your own data, it learns your gaps along with everything else. The people you've never served well don't show up in the training data, so they don't show up in the empathy either. It's the same trap as an untested persona, just faster (see principle 13, which is also Mike's).

What the tools tell you, and what they don't

On the digital empathy hub I wrote about watching a recording of someone searching a site for "bus shelter advertising", finding nothing and leaving. The analytics logged one session and a bounce. Watching it, you could feel them getting annoyed. The recording didn't do the empathy. I did, by sitting there and paying attention.

Melanie Spring told me she doesn't get enough hugs, and a typed one doesn't count. It's nice when someone writes "hug", she said, "but that's not actually as much as it might seem like, it doesn't really wrap me up with a hug." A lot of automated "personalisation" is the typed hug. She goes further in value human interaction.

Mike's agency, iPullRank, now spends a lot of its time on AI search, so this isn't a man who's afraid of the robots. He just knows what they're for.

If you want to try it, list every automated message a customer might get from you and read each one as if you'd just had a bad day. Then skip the sentiment dashboard for a week and read ten real emails or chat logs instead.

I used to sign off every episode of the show the same way, and Melanie's episode is a good place to hear it: "look after each other online because empathy is organic and you can't automate empathy." Bobby had a different plan. He suggested using "advanced AI to take over the motor functions of a company's engineers, using subliminal implants". One option anyway.

For the full conversation, listen to Mike King part two or read the written interview on kwasi.com. There's a longer take on the idea in what is digital empathy too.