Why "my job is too complex to automate" is Probably Wrong
The complexity of a job is no safeguard against automation. Only a subset of tasks must be automated to risk the job as a whole.
In 2018 I penned an op-ed in The Scientist titled “Rise of the Robot Radiologist.” It was a brief, somewhat sharp-elbowed review of automated medical image interpretation. The backlash from my radiologist colleagues and friends was swift and unhappy. Among a tirade of angry tweets and ad hominems one retort I heard again and again was: that’s not all we do. Recognizing white spots on a picture is not all we do. Recommending the right study is not all we do. Radiation dosing is not all we do.
This response reveals a fundamental misunderstanding of the dynamics of automation.
It’s time for a key take away: not every task must be automated to automate a role as a whole.
To understand this we first need to deconstruct work.
Deconstructing Work
A role can be broken down into multiple tasks. Tasks are the discrete functions that make up the role. A taxi driver picks up passengers, helps with their luggage, identifies the best route, makes small talk and so on.
Tasks are either Core or Ancillary:
Core tasks are those necessary and sufficient for the work. Without these tasks there is no job, with them you can get by.
Ancillary tasks may be important but are not absolutely essential.
For instance, a taxi driver’s core task is driving safely from A to B. If you can do this, you could be a taxi driver. Helping with your luggage, making small talk, and recommending eateries are ancillary tasks. They’re nice to have but unessential
Driverless cars can’t do everything a taxi driver does, but they don’t need to. Driverless cars pose a credible risk to taxi drivers because they can fulfill their core tasks. Only core tasks need to be automated to risk automating a role as a whole.
What about the ancillary tasks?
Broadly speaking they can be (a) done without, (b) done by less skilled labour, or (c) also eventually automated.
Task Dimensions
Let’s focus on Core Tasks for now. We can consider each Core Task along two dimensions: (1) Mental-Mechanical, and (2) Linear-Stochastic. Put differently, our tasks involve thinking and doing and these can be predictable and repetitive or unpredictable and dynamic.
Some tasks involve more of one than the other, but virtually all involve a little of each. A food packer in an assembly line has mostly Mechanical-Linear tasks, a poet has mostly Mental-Stochastic tasks, and an NBA player has mostly Mechanical-Stochastic tasks. This is an oversimplification but you get the point.
Automation Patterns
In the past, technology replaced Mechanical-Linear tasks. For instance, the stocking frame kicked out the weavers and industrial robots replaced food packers. When computers came around, Mental-Linear tasks were next. The computer spreadsheet, for example, bumped off countless accounting clerks.
Today, technology is graduating from Linear to Stochastic tasks. Here things are moving in reverse. Mental-Stochastic tasks are first in line. For instance, GPT-3 developed by OpenAI is an exceptional writer. Mechanical-Stochastic tasks are not far behind. Boston Dynamic’s dancing robots look more like fluid CGI with every press release.
We’re closer to building an artificial Shakespeare than a robot Michael Jordan but the gap is slim.
Summary
A job or a role is comprised of multiple tasks. Some tasks are essential for the role others are not. I call these Core and Ancillary Tasks respectively. Role automation is often dismissed as unfeasible because a role is “too complex”. This retort misses a crucial point: not every task need be automated to risk automating the role as a whole. Only automating Core Tasks is necessary and sufficient to risk a role. Ancillary Tasks can be done without, done by less skilled labour, or also eventually automated. Tasks can be described along a Mental-Mechanical and Linear-Stochastic matrix. Whereas historically automation was limited to Mental-Linear and Mechanical-Linear tasks, it is increasingly encroaching into Stochastic tasks.