AI Stops You Building Real Skills | Peak by Ericsson

AI makes it look like I can produce above-average output in any new skill. However, the question I kept avoiding was: Whether I had actually become better, or just faster at looking like it.

Over the past year, my worry going all in with AI was simple. I risk outsourcing my thinking to AI. I had been working to fix that, asking AI to question me rather than answer me outright. I thought that was enough.

It was not. That assumption was the real problem. Peak by K. Anders Ericsson showed me why.

Ericsson spent decades studying how experts become experts. His finding cuts against everything we assume about talent. Nobody is born an expert.

Our brain adapts to the right kind of pressure. People we call naturally gifted used that adaptability more than the rest of us.

The process has a name: deliberate practice. It is not repetition. Repetition gets you to an acceptable level and stops.

Deliberate practice is focused effort just beyond your current ability, with feedback, a clear goal, and full attention. Discomfort is the mechanism, not a side effect.

Ericsson’s framework is not universally accepted. The core mechanism holds well enough for me get started. If better evidence emerges, I will need to update myself.

The mechanism is mental representations. Every time you struggle with a problem, you build an internal model of that domain. Figure out why you failed and the model gets stronger.

An expert chess player does not calculate every move from scratch. They see patterns immediately. That pattern recognition is the product of thousands of hours of deliberate practice. My read is that Kahneman named this as System 1 thinking. Ericsson traces how it gets built.

The key word is struggle. You cannot build mental representations by watching someone else do the work.

Ericsson studied elite performers in structured domains, chess, violin, medicine. These fields have centuries of pedagogy and fixed definitions of what expert looks like. Knowledge-based skills are messier.

For Knowledge-based skills, the definition of expert keeps shifting as the field evolves. I am not building mental representations toward a fixed destination. I am building them while the field redefines what expert means.

This is where AI creates the trap. When I ask AI to produce marketing copy, the output is good. The struggle never happens. I never build the model.

If I already understand marketing deeply, AI helps me express that understanding better. But if I never built the foundation, I am producing output with nothing underneath it.

I ran an audit of my blog last year. I asked myself a simple question: Would this post help a potential customer decide? Most failed. They ranked well. They covered everything. They had none of my actual operational knowledge in them.

AI had done the construction work. The work was hollow. Posts with my operational knowledge in them started generating enquiries. The information-dense ones did not. Other things changed at the same time. I cannot isolate the method as the cause.

The fix is not to avoid AI. It is to change when and how you bring it in.

I now work through three stages when learning any new skill. This is my own framework, built from Ericsson’s principles and my existing process. It is not his.

The first stage is decide and diagnose. I map what I already know. I gather enough to make one decision:

• Is this skill worth my time or not.

I come out of Stage 1 with a clear yes or no. No AI yet.

The second stage is build the foundation. I gather sources, books, articles, and podcasts from people who know the domain. I load them into a single knowledge base and interrogate it. Right now I use NotebookLM for this. Five years from now a better tool may exist. The process is what matters, not the tool.

I ask three questions:

• What are the core mental models every expert in this field shares?
• Where do experts fundamentally disagree, and what is each side’s strongest argument?
• What ten questions would expose whether someone understands this subject or memorised facts?

Then I ask the tool to question me. Stage 2 is done when I can answer from my own understanding, not from looking things up. When the field moves, I return here before moving to Stage 3.

The third stage is deliberate practice with AI. Once I have the foundation, I bring AI in as a practice partner. Not to produce. To challenge.

I use it to find gaps in my reasoning, push back on weak arguments, and stress-test conclusions. The prompt I now use:

Respond as a Socratic teacher. Guide me through questions. Avoid direct answers. If I am wrong, say so. Do not agree to make me comfortable. Play devil’s advocate when my argument needs pressure testing.

This is slower than asking AI for answers. The slowness is the point.

Ericsson’s research shows that people who become good at something are not the ones with the most natural ability. They are the ones who did not skip the construction process. AI makes skipping it very easy. The output looks the same whether you skipped it or not. That is the trap.

I thought I had solved the AI problem by fixing my prompts. Peak showed me I had not. You can brief AI perfectly and still let it do the learning for you. The brief controls the output. Only the struggle builds the skill.