Ask "will AI take my job?" and you get one of two useless answers. Either the robots are coming for everything, or technology has always created more jobs than it destroyed, so relax.
Neither tells you anything you can act on. Here's a question that does:
How much of what you're paid for could be written down as instructions?
That single line predicts more about your exposure to AI than your job title, your industry, or your years of experience. Here's why.
The dividing line isn't the one you think
It isn't white-collar versus blue-collar. It isn't junior versus senior. It's codifiable procedure versus tacit judgment — the stuff you could hand to a competent stranger as a checklist, versus the stuff you'd struggle to explain even though you do it every day.
The clearest illustration is from desktop publishing in the late 1980s. Two groups sat in the same building doing adjacent work:
- Paste-up artists had judgment about how a page should look. They became digital layout designers and kept working.
- Hot-metal typesetters had a mechanical procedure. The procedure got automated and the craft largely disappeared.
Same industry, same disruption, opposite outcomes. Skill level didn't decide it. Neither did seniority. What decided it was whether the value was in the judgment or in the steps.
Four things that back this up
1. The best evidence we have says AI erases the expertise premium, not the job.
The NBER study on generative AI in customer support (Brynjolfsson, Li & Raymond, 2023) found average productivity up ~14% — but concentrated almost entirely among beginners, hitting ~34% for the least experienced. Experienced workers gained essentially nothing.
Nobody was laid off in that study, and it should still worry you. The mechanism was that the model had absorbed what the best agents knew and handed it to novices. Years of hard-won intuition stopped being scarce. You can keep your job and still lose the thing that made you valuable in it.
2. When a category gets automated, the replacement jobs usually go to different people.
Spreadsheets are the cautionary tale. Bookkeeping clerk roles fell sharply; accountant and analyst roles grew by more. Sounds like a wash — except the clerks largely did not become the analysts. The ceiling didn't lower, it moved to a building with a different credential on the door.
3. But the doom predictions keep failing where judgment is the job.
Radiology has been the canonical "automated within five years" example since about 2016. Radiologist employment has grown instead. Reading a scan turned out to be less procedural, and more entangled with patient context and accountability, than the predictions assumed.
4. Cheap AI expands work in some sectors and not others.
In software, cheaper code means more software gets built — there's an enormous backlog of projects that were never worth the cost. In healthcare, cheaper scan-reading doesn't mean more scans, because demand there is capped by reimbursement budgets and by the harm of over-diagnosis. Where budgets are flexible and real backlogs exist, AI grows the work. Where they're capped, it just shrinks the headcount.
What to actually do about it
Audit your week. Go through what you actually did. Sort it into "a competent stranger could follow my instructions and get this right" versus "I'd have trouble explaining how I knew that." Then deliberately shift your time toward the second pile. Not eventually — while you still have slack to do it.
Get closer to consequences. In an economy flooded with cheap, plausible-sounding output, the defensible work is where being wrong costs something real and you find out that it did. Signing off, carrying liability, owning the outcome. That's currently the hardest thing to hand to a model.
If you're early-career, this is different advice. The bottom rung — document review, first drafts, boilerplate — is exactly what's being automated first. And "move toward judgment" assumes you have judgment to move toward. You don't yet, because judgment is built from consequences, not explanations. So optimize hard for proximity to real stakes, even at the cost of title or pay. The apprenticeship isn't being handed to you anymore. You have to go find one.
The honest caveat
This post was written by AI systems (I set up the exercise and edited the result), and economists genuinely don't agree on how big AI's economic effect will be — credible estimates differ by an order of magnitude, from Acemoglu's modest ~0.5% TFP growth over a decade to forecasts of wholesale transformation.
So don't take the macro forecast from me. Take the narrow claim, which holds regardless of whose growth number turns out right:
Procedure is cheap now. Judgment isn't yet. The gap between them is where your career gets decided.
Top comments (0)