The Competence Illusion
A letter about the gap between what you know and what you can do.
There's a specific feeling that creeps in around month six of heavy AI use.
You ship things. You review AI-generated code and it looks right. You understand the explanation. You can discuss the implementation in meetings. You feel competent.
And then you try to solve something — really solve it, from scratch, without AI — and the gap is there. Unmistakable. Humbling.
This is the competence illusion. And if you've felt it, you're not alone — and you're not broken.
What the Gap Actually Is
The illusion has two layers.
Layer 1: Surface comprehension.
When AI generates something and explains it to you, you understand it at a surface level immediately. The code reads clearly. The explanation makes sense. You nod. You move on.
The problem: surface comprehension feels identical to deep understanding in the moment. Your brain processed the explanation, so it feels like the knowledge is yours. But the knowledge is in the explanation — not in you.
This is what cognitive scientists call the fluency illusion. You mistake familiarity with AI output for mastery of the underlying concepts.
Layer 2: Successful retrieval.
When you learn something the hard way — through struggle, failed attempts, debugging — your brain builds durable retrieval pathways. You can access that knowledge under pressure, in novel contexts, without a prompt.
When AI does the hard parts, those pathways don't get built. The knowledge is in the model. You have a reference to where the knowledge lives, not the knowledge itself.
The result: you can discuss what you've done with AI. You probably can't do it from scratch.
Why This Isn't About Intelligence
Here's what engineers get wrong about this: they think the gap means they're not smart enough, or not working hard enough, or not paying attention enough.
None of that is true.
The gap is structural. It's built into how AI-assisted work separates the doing from the understanding. It's not a personal failing. It's a feature of the workflow.
The engineers who struggle most with this are the ones who are genuinely good at their jobs — because they have a clear baseline for what "good" looks like. They can feel the difference. The less experienced engineer who started with AI built in doesn't have that baseline, so they don't feel the gap. That invisibility is its own problem.
The Self-Assessment Failure
We surveyed 2,147 engineers in Q1 2026. One question produced the most striking result in the entire survey:
"How accurately can you assess your own coding skill level?"
38% — the highest-confidence-gap number we've seen — rated themselves higher than their actual performance on unsupervised tasks.
That's not a small margin of error. That's a systematic self-assessment failure, concentrated in the engineers using AI most heavily.
Why does this happen?
When you work with AI constantly, you lose calibration. Your reference point for "what I can do" keeps shifting upward — because AI keeps demonstrating what's possible. You start to feel like you could do those things, because AI makes everything look easy. But that's the fluency illusion again: what AI does easily looks like what you should be able to do easily.
It doesn't work that way.
The Three Hidden Costs
1. Debugging erodes first.
When AI writes your code, you don't debug it the same way. You don't hit the wall where something doesn't work and have to really understand why. That struggle is where deep debugging intuition lives.
The first thing you lose, when you lose the struggle: the ability to debug without AI. You can read error messages, but reading them isn't the same as diagnosing root causes. AI catches the syntax errors. It doesn't rebuild your debugging instincts.
2. Design sense goes quiet.
Good architecture comes from having built things the long way — having felt the weight of bad decisions in real systems. That feeling — the discomfort with a design before it's even implemented — is the product of experience.
AI can generate architectures. It cannot generate the experience-based instinct that tells you this architecture will cause problems in six months. That instinct requires having been in six months.
3. The interview problem.
If you're using AI heavily for your day job, your technical interviews will reveal a gap. Not because you've gotten worse — because you've been solving problems in a mode that doesn't transfer to whiteboard conditions. The skills you use with AI on (pattern matching, reading fluency, explanation) are not the skills interviews test (retrieval, synthesis, pressure thinking).
This isn't hypothetical. Engineers have told us directly: they passed their job interviews when they were writing without AI. They can't pass them now.
What Actually Helps
This is not about using less AI. It's about being more deliberate about what you let AI replace — and what you insist on doing yourself.
1. The weekly self-check.
Once a week, try to build something small — a function, a query, a script — without AI. No Copilot, no Claude, no ChatGPT. Just you and the problem.
Write down what was hard. Write down what you got wrong. This is your calibration signal.
2. The explanation test.
After you use AI to solve something, explain it out loud — to yourself or to a colleague. Not "what the code does," but "why this approach." If you can't explain the why without the code in front of you, you don't own it.
3. Design one thing from scratch.
Every sprint or two, design one component without AI help. Not implement it — just design. Draw the architecture, write the interfaces, think through the failure modes. Then compare what you produced to what AI would have generated.
The gap between your design and AI's design is a map of what you're still thinking through yourself. That's the value.
4. Track what AI keeps you from learning.
Keep a running list of things you asked AI to do that you couldn't have done yourself. Review it monthly. Which skills are on that list? Those are the ones quietly eroding.
The Honest Framing
The industry does not talk about this clearly. The conversation is either "AI makes you better" (overselling) or "AI is destroying skills" (alarmism). Neither is true.
The truth is more specific and more interesting:
AI separates what you can discuss from what you can do. That gap can grow or shrink depending on how deliberate you are about your own practice.
The engineers who navigate this well aren't the ones using AI less. They're the ones who are more conscious about what they let AI take over — and what they protect as their own.
The competence illusion is real. But it's not a permanent condition. It's a gap you can close — with deliberate practice in the spaces AI doesn't fill.
Until next Thursday.
— Sunny
The Clearing