Essay
The competence trap
AI can make us dramatically more productive while quietly weakening the understanding beneath the output.
-
01
Leverage
The artifact improves faster than the underlying skill.
-
02
Boundary
You must know where your own understanding ends.
-
03
Practice
Some resistance must be preserved on purpose.
AI should make us more capable, not only more productive.
Borrowed competence
GPS can make you excellent at arriving while teaching you almost nothing about the city. AI creates the same effect for intellectual work. Someone who could not build an application can now create one; a weak writer can produce a polished email; a founder who has never modeled a business can request a spreadsheet. The output improves. The underlying competence may not.
This is the competence trap. It matters because leverage is attractive, especially to entrepreneurs. But leverage carries a hidden question: when the tool produces something wrong, do you know enough to notice?
The artifact is weaker evidence
Productivity and competence used to be tightly connected. Good software suggested understanding of software. A clear memo suggested understanding of the argument. AI breaks that connection. A convincing artifact can now be produced with only a surface-level model of what it contains.
The danger grows with consequence. Mediocre marketing copy is usually recoverable. A database migration that corrupts customer data, a contract built on the wrong jurisdiction, or a financial model that classifies refunds incorrectly is different. Borrowed competence becomes expensive when failure is difficult to reverse.
Know which level you are on
There are three useful levels. At the first, you can do the work yourself. At the second, you cannot perform every detail, but you understand the system well enough to judge good work from bad. At the third, you can only judge whether the result looks convincing.
AI makes the third level extremely productive and extremely dangerous. One of the most valuable founder habits will be asking: am I evaluating this by substance, or only by surface?
Production mode and training mode
The answer is not to avoid AI. Separate production mode from training mode. In production, use leverage: automate, generate, and delegate. In training, deliberately keep some resistance. Write the first version yourself. Debug one problem before asking the model. Read the original source. Build one financial model from a blank sheet. Explain the architecture to another person without opening the chat.
The strongest people will use AI on top of competence, not instead of it. Small teams can gain extraordinary leverage, but they also have less redundancy. If nobody retains a human model of the critical system, independence from headcount can become dependence on the tool.