Software Update

We taught people to be machines because machines couldn’t be people. Then, they could.

For most of the last century, a good job meant executing a procedure reliably. Autor, Levy and Murnane measured the turn: between 1970 and 1998, routine cognitive work drained out of the American economy as computers arrived to do it. Before that, the safest career available was being a slightly worse computer. Reconcile the ledger. Check the form. Pull the numbers, put them in the right cells, don’t make errors. We built schools to produce this, tests to certify it, and entry-level jobs to train it.

Then software got good at it, and we all started saying the obvious-sounding thing: fine, now the human skills matter. Empathy. Connection. Relationships. I said it too. My first version of this argument was that being good with people would beat being good at the task.

I don’t think that’s true anymore, and the evidence against it is unpleasant.

In 2023, licensed clinicians read 195 real patient questions, each answered twice: once by a doctor, once by a chatbot. They rated the chatbot’s answer empathetic 45% of the time. The doctors’: 4.6%. In a preregistered debate experiment, GPT-4 given a few demographic details out-persuaded human opponents 64% of the time, winning on precisely the thing we call irreducibly human: reading the person in front of you. Eight minutes of conversation with a language model cut entrenched conspiracy beliefs by about a fifth, and the effect held two months later. An AI companion relieved loneliness roughly as well as talking to an actual person; the mechanism the researchers isolated was “feeling heard.”

Meanwhile the jobs vanishing are the people-facing ones. Retail sales shed 850,000 positions in a decade. Front-office administrative work fell from 5.8% of American jobs in 1990 to 2.2% today. The Bureau of Labor Statistics expects customer service, one of the largest occupations whose entire job description is deal with a human being, to lose another 142,000 by 2035.

So “be more human” cannot mean “be warmer.” Warmth is now a commodity: instant, unlimited, and rated higher than ours by the people trained to judge it.

What’s left is less flattering and much harder. Machines already know more than we do, which rules out knowledge. Judgment remains - the capacity to assess a risk well: knowing what to focus on when the trade-offs are real, this money and these months spent here and not there, with your name on the outcome. A machine can’t take a risk. It has no desires, owes nothing, needs nothing; it just is. Humans take risks, and being the person others will allow to take them is the job.

Which makes what companies are currently doing look very strange.

Firms say they want judgment. They screen for procedural competence: the Excel test, the case interview with a known answer, the years-of-experience filter. And at the same moment, they are deleting the jobs where judgment used to be grown. Stanford’s payroll data, covering millions of workers tracked monthly, shows 22-to-25-year-olds in the most AI-exposed occupations running about 19% below where they should be, almost entirely through reduced hiring rather than layoffs. The researchers’ own hypothesis: AI may be automating “the checkable, process-intensive tasks that historically justified entry-level headcount.”

Those tasks were tuition. You learned judgment by making small consequential decisions badly, under supervision, for a few years, in a job that mostly consisted of clicking. That is how a company hedges risk: it hands you the low-risk, low-reward calls first, and the intuition you build there is what earns you the ones that pay. Nobody is teaching that now. We automated the apprenticeship and kept the entrance exam.

Architecture already ran this experiment, and it is worth looking at closely.

Before CAD, becoming an architect meant years at a drafting board. You inked someone else’s details, over and over, at a scale that forced you to understand what you were drawing. You could not draft a wall section without learning what goes inside a wall: the flashing, the vapor barrier, the order in which the trades arrive. The drudgery was the curriculum. Nobody designed it that way. It was simply that the only way to produce a set of construction documents was for a young person to spend three years drawing them.

AutoCAD shipped in 1982, and within a generation the boards were gone. The Bureau of Labor Statistics is blunt about where the work went: CAD and building information modeling “increase drafter productivity and allow engineers and architects to perform many tasks that used to be done by drafters.” Drafting employment is projected to grow 1% over the coming decade, which is flat in an expanding economy. The task did not disappear. It migrated upward, to the person who used to learn the job by doing it.

Here is the part I find remarkable. Architecture noticed. It is one of the few professions that responded to losing its apprenticeship by legislating a replacement. To be licensed in the United States you must still log 3,740 hours of supervised experience, and the single largest required category, 1,520 hours and more than a third of the total, is project development and documentation. The profession automated the drafting and then made the drafting hours mandatory anyway, because what the drafting taught turned out to be the job. The median path to licensure still runs 12.3 years, and that is the fastest it has been in a decade.

Every other industry is doing the first half of that and skipping the second.

Nineteen expert endoscopists, thousands of procedures each, were given AI assistance. Afterward, their detection rate on unassisted procedures fell from 28.4% to 22.4%. A fifth of their skill, gone, in months. Judgment is a practice, and a practice atrophies when you stop doing it.

Drawing a building slowly changes the building, because the medium answers back. A line you have to commit ink to is a line you consider. A detail you redraw at three scales is a detail you find out is wrong at the second one. Donald Schön, who sat in architecture studios to work out how professionals actually think, described design as a reflective conversation with the situation: you shape the thing, the thing talks back, and you revise what you were trying to do in the first place.

Notice what that implies. The goal is one of the things being discovered. This is precisely what optimization cannot do, by construction. To optimize you need a stated objective, and the creative process is the procedure by which you find out what the objective should have been. A system that returns a hundred options against a brief is fast, and it is doing something categorically different from an architect who draws a wall section and realizes the whole scheme wants to be one bay narrower.

The process leaves marks on the result. A building that was drawn slowly carries decisions no one could prove were the right ones: proportions that are what they are because someone spent four months with the problem and the problem pushed back. Those four months are how you get this building instead of a correct answer, something worth having for what it is, beyond the boxes it checks. We have precise language for whether a design meets its brief and almost none for whether it is that thing and no other.

So the comfortable version of this essay, that machines take the drudgery and humans get the meaningful part, has it backwards. The machine-like discipline was the practice that crafted the human part. Being more human is a harder job than the one we just deleted. Architecture worked that out and made itself slower on purpose. The rest of us deleted the drudgery, kept the entrance exam, and called it progress.