Jab Petersen / AI Won’t Save You

Chapter 5 · The Apprenticeship Paradox

How do beginners become experts when AI does the beginner's work?

Two paths can produce the same polished result. Only one of them produces a person who can tell when the tool is wrong.

Here is the question underneath this chapter: if technology performs more of the work through which beginners used to become experts, how do beginners become experts?

Two paths that look the same

The traditional path to expertise looks like this:

BEGINNER → WORK → MISTAKES → FEEDBACK → PATTERN RECOGNITION → JUDGMENT → EXPERT

AI creates a second path:

BEGINNER → AI → EXPERT-LOOKING OUTPUT

From the outside the results can look identical: a polished report, working code, a sophisticated essay. But similar artifacts do not guarantee similar people. The hidden question is what capability exists when the tool is uncertain, unavailable or confidently wrong.

Dead friction and productive friction

Not all difficulty is valuable. Typing a manuscript on a typewriter was harder than on a laptop, and it did not produce better thinking. Copying data between systems is exhausting and teaches nothing. Call that dead friction, and remove it.

Then there is productive friction: explaining an idea before seeing the answer, debugging long enough to understand the architecture, talking to a customer before asking AI to simulate one. Learning researchers call versions of this "desirable difficulties" (Robert and Elizabeth Bjork) and "productive failure" (Manu Kapur). It hurts a little because learning needs prediction plus correction.

I spent roughly eighty hours on my first five-minute film. Some of those hours were dead friction I would delegate gladly today. Others built the eye that now lets me spot a weak cut or an unusable AI clip. If a model had made the whole film instantly, I might have had a better video and become a worse editor.

Decide: performance or practice?

Before you open the tool, decide which mode you are in.

  • Performance mode: the result is the priority. The deck is due tomorrow, the bug must be fixed. Use the strongest tools you have.
  • Practice mode: your development is part of the output. You are learning a language, to code, to write, to negotiate.

In practice mode, keep some of the struggle on purpose. Tell the AI:

  • Do not solve this. Give me one hint.
  • Ask me questions until I find the error.
  • Critique my attempt without rewriting it.
  • Make me teach this back to you.

The same technology that can bypass apprenticeship can become an apprenticeship machine.

The practice sandwich

A simple learning design with effort on both sides of the machine:

  1. Before AI, predict. What do you think the answer is? What would you try first?
  2. With AI, compare. Ask for explanation, critique, alternatives and hints.
  3. After AI, reconstruct. Close the answer. Explain the principle from memory. Apply it to a new case.

It works especially well for languages, programming, math, writing and interview preparation.

Watch your learning debt

Engineers talk about technical debt: shortcuts that speed things up now and cost more later. AI creates learning debt. You ship faster by accepting explanations you do not understand. When the system fails in an unfamiliar way, the missing understanding comes due.

Some learning debt is rational. You do not need transistor physics to use a laptop. But keep it visible:

What am I choosing not to learn because the tool can handle it? Under what conditions would that ignorance become expensive?

Earn the right to automate

A practical loop: attempt, amplify, explain, apply, offload. Make a first attempt. Use AI for hints and critique. Explain the answer without copying it. Apply it somewhere new, because transfer is the real test. Then, once the skill is yours, decide what no longer deserves your attention.

The goal is not to stay a beginner forever out of fear of automation. It is to earn the right to automate.

Don't automate away an experience that is still teaching you something you need.