Part II · Chapter 3 · Human Capital · lesson 6
Which skills decay, and how fast
12 min
Builds on Why some things get cheaper and some never do
Part II · Chapter 3 · Human Capital · lesson 6
12 min
Builds on Why some things get cheaper and some never do
Chess is one of the hardest things people do. Machines have been better at it than any human for decades.
Loading a dishwasher in an unfamiliar kitchen is something most nine-year-olds manage. No machine does it reliably.
Whatever separates those two, it is not difficulty. Difficulty is the intuition everyone brings to this question and it points in the wrong direction almost every time.
Predict
Exposure to automation tracks how specifiable a task is, not how difficult it is.
A task is specifiable to the extent that its inputs can be enumerated, its allowed actions written down, and its success checked without a human deciding. Chess is completely specifiable and extremely hard. Settling a distressed child is not specifiable at all and is something most people can do.
Difficulty and specifiability are close to independent, and only one of them predicts anything.
Specifiability is the largest of them and it is not alone. Four properties, taken together, do most of the work.
Specifiability. Can the task be stated as inputs, permitted actions and a checkable success condition? If a competent person could write down the rules and a stranger could follow them, the answer is yes.
Recorded examples. Does a large body of worked cases exist, with the answers attached? Where an industry has been writing down what it did and whether it worked for thirty years, the raw material is already assembled. Where the knowledge lives in people's hands, it is not.
Environmental variability. How much does the situation differ from instance to instance, and how much of the difference matters? A production line is a controlled environment. A stranger's house is not.
Cost of error. How bad is the worst mistake, and can it be caught before it does damage? Where errors are cheap and reversible, an imperfect system is adopted quickly. Where they are catastrophic and irreversible, a system has to be far better than a person before anyone will use it, and someone still has to be accountable.
Not the same thing
Takes long training, sustained concentration, or a rare aptitude to perform.
Cannot be reduced to enumerable inputs, permitted actions and a checkable result.
Which is which? Put each one on a side.
Grading multiple-choice examinations at scale
Deciding which of two candidate explanations for a failure to investigate first
Detecting a specific pattern in a large volume of images with known labels
The usual framing is that jobs disappear. That happens, and it is less common than the thing that actually happens, which is that the composition of a job changes and the scarce part moves.
A job is a bundle of tasks. Automation takes the specifiable tasks out of the bundle first, because those are the cheap ones to take. What is left is a smaller bundle, more concentrated in whatever was unspecifiable, and the wage attached to it depends on what the remaining tasks are worth and how many people can do them.
Two very different outcomes follow from that, and which one you get depends on the residue.
If the specifiable part was the barrier to entry — the reason few people could do the job — removing it enlarges the pool and pay falls. Anyone can now do most of it.
If the specifiable part was the drudgery and the unspecifiable part was the value, removing it raises output per person and pay can rise. The same person now spends their day on the part that was always scarce.
Check
Instead of asking whether a job will be automated, which nobody can answer, the tractable question is which parts of it are specifiable and what would be left.
Take any role and split it into tasks. For each one: could the rules be written down, do recorded examples exist, does the situation vary, and how expensive is a mistake? The tasks that score badly on all four are the exposed ones, and they will go first regardless of what anyone predicts about the field as a whole.
Then ask the second question, which is the one that decides the wage: of what remains, how much was the reason the pool was small?
Practitioners consistently overestimate how unspecifiable their own work is, and the reason is structural rather than vanity.
Expertise is largely the process of turning explicit rules into automatic behaviour. A skilled person has stopped noticing the rules they follow, so when asked whether their work can be written down, they honestly report that it cannot — while a novice in the same field, who is still consciously following the steps, would give a different answer.
The practical consequence is that the best judge of how specifiable a task is tends to be someone who learned it recently, or someone who has had to train another person from nothing. Both of those people have seen the rules written down, because they had no other way to proceed.