Teach the Apprentice, Starve the Master? What AI Is Built to Copy Decides Your Job

Teach the Apprentice, Starve the Master? What AI Is Built to Copy Decides Your Job
Power Behind the Screen

Case of the Week, Min Wu, PhD · ai-public-health.com

In October 2026, I attended a virtual seminar on AI, and it covered a lot of ground: the workforce, innovation, regional competitiveness. The students in the chat cared about one of those themes far more than the others. One question read: I am an art major — what will AI do to the art job market in the coming years? Others followed in the same shape. The major changed, but the question didn't. They weren't asking about innovation or regional competitiveness. They were asking whether the job they are training for will still be there when the training is done.

Reading that chat, I kept hearing an old Chinese saying: 教会徒弟,饿死师傅 — teach the apprentice everything, and the master starves. For most of history it was a warning about people. Watching those questions scroll by, it began to sound, to me, like a description of some of the AI we are building, and I wanted to give those students a better answer than a chat box can hold.

Why the Students' Question Makes Sense

Chapter 1 of my textbook, Artificial Intelligence in Public Health, explores AI from four angles: technology-based, role-based, concept-based, and ethical. The concept-based definition describes AI as a machine built to mimic human cognitive functions to understand the world. Read that way, the students' worry is understandable. Every function the machine learns from us — reading, remembering, classifying, deciding — is one more lesson the apprentice has absorbed. Learn enough of them in the right bundle, and the bundle can start to look like a job.

But look closely at what the definition says AI copies. Functions, not roles. A function is a single mental task, such as reading a form or recognizing a pattern. A role is something else: a bundle of functions arranged around what one human body and one human workday can hold. A scheduler, a coder, a case manager — each is a particular arrangement of perceiving, remembering, and judging, sized for a person. The machine learns the parts. What it builds with the parts is a design choice. That choice, I believe, may matter as much as the learning itself for what happens to the master.

I don't want to promise more than I know. Some jobs, or parts of jobs, will likely change or shrink. Some AI is being built to copy tasks that existing roles perform, and where the copy works well enough, those roles may need fewer people. For that kind of AI, the old saying may hold.

Two Apprentices

In an earlier post, Superhuman vs. Non-Human, I split AI strengths into two kinds. The split grows out of Chapter 3 of my textbook, on technology-based design. Superhuman strengths do a familiar job better: faster, tireless, with perfect recall. Non-human strengths do something no human role was ever built to do, such as selective forgetting (letting go of outdated information on purpose), or holding many data streams at once and noticing a pattern among them. That post was about design. This one returns to the same split to ask the seminar students' question: what happens to the job? In plain terms, one apprentice copies the master; the other complements the master.

The superhuman apprentice is the one the old saying warns about. It reassembles the learned functions back into the master's exact role — the same tasks, in the same order, just without fatigue, sick days, or a salary. When it works well enough, little in the workflow needs to change except who does the work. That is the path toward replacement, and it is the path the students worry about.

The non-human apprentice learns the same functions but puts them together into a shape the master never had. Take the home-care robot from that earlier post. Built as a superhuman, it performs the home health aide's daily check-in — same questions, same schedule, never a forgotten follow-up. Built as a non-human, it combines how a person walks, sleeps, eats, and sounds into one signal, and one day it flags a change no checklist was designed to catch. Now someone has to read that flag, judge whether it matters, and decide whether to call the family or the clinic. That work may not have existed before, because no one could produce that signal before. In this example, the apprentice did not take a chair. It may have added one.

Economists have measured something like this across a whole labor market. David Autor and colleagues traced new job titles in the United States over eight decades and found that most U.S. employment by 2018 was in job specialties introduced after 1940. They also found that new work tends to emerge from technologies that help workers do more, while technologies that take over workers' tasks slow new work down (doi:10.1093/qje/qjae008). Erik Brynjolfsson calls one risk the Turing Trap: too much focus on human-like AI that replaces workers rather than helping them (doi:10.1162/DAED_a_01915). I came to a similar split from the design side, through a home-care robot; they came to it through economics. That gives me some confidence, but only some.

The history these economists studied was mostly about fixed tools. A typewriter or a spreadsheet does the same thing on its last day as on its first, and people built new work around it. AI is different: it keeps learning, and it can adapt to the very people working beside it. The new work that grows around AI may itself become something AI learns, so a chair it adds today could be a chair it learns to fill tomorrow. I don't know how much this changes the historical pattern, but I would be careful about assuming it holds unchanged.

What Complementary Can and Cannot Do

So is the answer simply to build complementary AI and let the jobs follow? I want to say yes, and I can't, for two reasons.

First, a complement can still shrink a team. If a tool lets one analyst do the work of two, and the amount of work stays the same, the organization needs fewer analysts — even though no single task was handed to a machine. Whether new jobs appear depends on whether the new capability creates new demand, and that is not something the designer controls.

Second, design intent does not travel with the product. A system built to assist a nurse can be purchased to replace one. A designer can build a tool to help; the decision about how it is used is made by someone else, often in a budget meeting the designer never attends.

What I can say honestly is narrower. AI built to mimic a role is aimed at a chair that already has someone in it. AI built to do what no role does is aimed at empty space. The second kind seems more likely to create new work. Neither kind guarantees it, and I expect the real picture in most fields to be a mix of both.

What I Would Tell the Student in the Chat Box

The question "what will AI do to my major's job market?" is hard to answer because a major is not one job. It is a bundle of tasks, and AI will treat each part differently. I can't answer it for any single major. But here is a pair of questions you can ask about your own field:

Which part of the work in my field is AI being built to copy?

Which part of AI's output will need a person who understands both the machine and the field to read it, check it, and act on it?

The first part is where the apprentice is learning the master's job. The second part is where new chairs may appear — the person who reads the robot's flag, who audits what a model got confidently wrong, who decides what a system should be allowed to forget. That second kind of work seems likely to reward what students are in school to build: deep knowledge of a field, not speed at its routine tasks. If I were a student today, I would lean toward it.

This holds even for the art student. Image generators are about as close as AI comes to copying a role's finished product. What they do not do is decide what is worth making, for whom, and whether the machine's version is right. That judgment, at least for now, still sits with the art student.

Whatever your major, you likely have some time to make that shift. In health informatics, I watched the move from paper records to electronic health records take many years, slowed by regulation, privacy rules, cost, and the slow work of earning trust. AI may move faster than that in some fields and slower in others. But in most fields, adoption depends on far more than what the technology can do, and that gap gives students time to choose where to stand.

The Observer's Insight

The old saying assumes the apprentice wants the master's job. For most of history that was true, because there was only one job to want. A machine that learns our functions does not want anything. It goes largely where its designers, and the organizations that adopt it, point it. Point it at an existing role, and it copies that role; the saying may come true, one task or one job at a time. Point it at what no role was ever able to do, and it complements the people around it; the apprentice need not starve the master; it may instead give the master something new to master. Which way it points is still being decided, partly by the people who sit in seminars like that one, typing their questions into the chat.

Copying takes the job; complementing makes one.


Affiliate notice: As a member of the Springer Nature Author Affiliate Program, I may earn a small commission from purchases made through this affiliated link to my book. Support the author by checking out my textbook, Artificial Intelligence in Public Health: https://tidd.ly/4mH9389