Jab Petersen / AI Won’t Save You

Chapter 3 · Don't Outsource the Part That Matters

Outsource production. Never outsource ownership.

AI does not have to make you think less. But thinking moves, and if you outsource both the work and the judgment, all that is left is clicking approve.

Every meaningful task produces two outputs. Writing produces words and, quietly, clearer thinking. Programming produces software and an instinct for debugging. Selling produces revenue and an understanding of who your customer really is.

Call the first performance. Call the second formation.

AI can raise performance much faster than formation. Often that is fine. Nobody needs to become a typographer to produce a readable document. But when performance rises far above formation, you depend on output you cannot properly judge. That is the performance trap.

Thinking does not disappear. It moves.

In 2025, Microsoft researchers surveyed 319 knowledge workers about 936 real examples of using generative AI at work. Higher confidence in AI went with less reported critical-thinking effort. Higher self-confidence went with more. The researchers also found that AI shifted thinking toward verification, integrating responses and stewarding the task.

A calculator removes arithmetic and makes choosing the right equation more important. Search removes some retrieval and makes judging sources more important. AI removes drafting and makes specification, verification and judgment more important.

That only helps if you actually do the new thinking. If you hand over production and judgment, nothing remains except approval, and approval is dangerously easy when the output looks polished.

Machines replaced most physical labor, yet the people who still go to the gym are strong. Thinking will go the same way. The people who do well will not be the ones who lean back and let themselves be served. They will be the ones who use AI as a sparring partner.

Before the Prompt

For any important problem, spend three minutes writing two sentences before you ask anything:

My current answer is ______ because ______. I would change my mind if ______.

Do not make it beautiful. The point is to create a position the machine can push against. Then give it an adversarial role:

  • What am I assuming?
  • Give me the strongest case that I am wrong.
  • What evidence would change the conclusion?
  • What would a smart person who disagrees say?

A gym does not make you stronger by moving the weight for you. It gives you something to push against.

Keep the loop closed

The working architecture is simple:

YOU → AI → YOU → REALITY → YOU

The human appears before the AI, after the AI and after reality. Written as steps: think, amplify, decide, ship, observe, repeat. This is not a sentimental defense of human superiority. It is an accountability structure. The human owns the goal, the standard, the decision and the consequence.

Three seats at the table

For important work, picture three seats.

  1. Creator: what are we trying to make true?
  2. Critic: what is weak, missing, false or self-deceptive?
  3. Owner: what will we actually do, and who accepts the consequence?

AI can sit in the first seat as a creative partner. It is excellent in the second. The third must stay clearly assigned to a person. The danger starts when responsibility gets foggy because "the AI recommended it." A recommendation is not an owner.

In 2023 a New York court sanctioned lawyers who had filed a brief citing court decisions that did not exist; ChatGPT had invented them. The tool wrote the draft. The lawyers' names were on the filing.

Set a verification budget

Checking every comma removes the leverage. Checking nothing is gambling. Match verification to consequence with two questions: How bad is it if this is wrong? Can I undo it?

  • Small and reversible, like a brainstorm or an internal draft: a glance is enough.
  • Public or costly, like a factual claim or a financial figure: check every fact and number that matters.
  • Irreversible, or touching other people's health, money or rights: full review, or a qualified human.

AI's fluency pushes you toward one style for everything: glance, approve, move on. Strong operators do not work that way.

The Stranger Test

Ever asked AI to fix something it never had the context for? Before you send an important prompt, ask: if I had to give this exact job to a stranger, what would they need to know? Two extra sentences can save an hour of rework.

The model can be the author of a draft. It cannot be the owner of a consequence.