The experiment nobody warned me I needed to run.
I was writing a three-sentence status email.
Not a system design doc. Not a complex PR description. A status email.
And I caught myself opening ChatGPT before I’d finished the second sentence.
That’s when I realized something had gone wrong. Not with the tool with me.
So I ran an experiment. 30 days. No ChatGPT. No AI writing assist. No Copilot. Just me, a keyboard, and whatever was left of the brain I’d spent years building and then quietly outsourced to a server farm.
What happened was uncomfortable. Then revealing. Then honestly kind of obvious in hindsight.
TL;DR: AI didn’t make me stupid. But I’d made thinking optional, and my brain had taken the offer. This is what 30 days of going cold turkey actually looked like and what I changed on the other side.
The dependency I didn’t know I’d built
There’s a concept called cognitive offloading using external tools to handle mental tasks your brain would otherwise process.
Writing a grocery list is cognitive offloading. Setting a calendar reminder is too. These are fine. Great, even.
The problem is when offloading becomes total.
Researchers at UC Santa Barbara found that when people consistently outsource a cognitive task to an external system, their ability to perform that task independently degrades. The neural pathway responsible gets less activation. Less reinforcement. It doesn’t disappear it just gets harder to access.
For developers, this happened in a very specific sequence.
It started with boilerplate. Fine. Then debugging shortcuts. Reasonable. Then architecture brainstorming. Then documentation. Then PR descriptions. Then email.
Then thinking.
The old Stack Overflow workflow had friction built in. You had to read five answers, reconcile conflicting opinions, understand your problem well enough to evaluate whether a solution applied. That friction was annoying.
It was also where the thinking happened.
ChatGPT removed the friction. Which felt like an upgrade until the friction turned out to be the feature.
The tell for me was a service boundary design session. Twenty minutes in, I hadn’t made a single decision. I’d been refining my prompt.
I wasn’t thinking about the problem. I was thinking about how to describe the problem to the AI.
Those are not the same skill. One builds domain knowledge. The other builds prompt fluency. Both useful. But I’d let one replace the other without noticing.
The 30-day log
I want to be clear: I didn’t do this to prove AI is bad. I did it because I’d noticed a specific, measurable decline in my ability to think independently.
Here’s what actually happened.
Week 1: the reach
Days one through three were fine. Then the reach started.
Every time I hit cognitive friction mid-paragraph, mid-design, mid-debugging session my hand moved toward a new tab. Before my brain had even registered the obstacle.
I started counting. Day five: eleven reaches. Eleven times I caught myself about to outsource a thought I hadn’t finished having yet.
The worst moment was a strategy doc for a service migration. I stared at a blank page for twenty minutes. Not because I didn’t know the answer. Because I’d lost the tolerance for not knowing it immediately.
Week 2: the grind
Everything took longer. PR descriptions that took three minutes took ten.
But around day ten something shifted. I started actually reading error messages. Not skimming for keywords to paste into a prompt. Reading them.
I caught a subtle race condition on day twelve that I’m fairly certain I would have missed with AI assist on. The AI would have pattern-matched to the surface bug. I would have shipped the fix without finding the underlying concurrency issue.
Small thing. But it stuck.
Week 3: the return
Around week three, something comes back.
It doesn’t feel like a superpower. It feels like a muscle waking up — that slightly uncomfortable awareness of something starting to work again.
Ideas started arriving differently. Not from prompts. From the background hum of just thinking.
I’d be mid-run, not trying to solve anything, and a clean solution to a two-day-old problem would just surface. That used to happen to me constantly. Somewhere in the past year it had mostly stopped.
Turns out I’d been interrupting the incubation phase every time I reached for the prompt box.
Week 4: clarity
By day twenty-five I wasn’t faster than I’d been with AI assistance. The tools are genuinely faster for certain tasks.
But I was thinking differently. Problems felt three-dimensional again. I was holding more context simultaneously, making more lateral connections, writing code that reflected actual decisions rather than synthesized suggestions.
Day thirty: a 1,200-word technical design doc, just under an hour, no prompts. The coffee went cold because I forgot it was there.
That hadn’t happened in a while.
What AI is actually doing to your brain
This isn’t a hot take. This is neuroscience.
When you perform a cognitively demanding task repeatedly, the neural pathways responsible for that task undergo myelination the process by which neurons become faster and more efficient. This is how expertise is built. Repetition, resistance, reinforcement.
When you consistently offload that task, those pathways get less activation. The myelin doesn’t build. The skill stays effortful instead of becoming automatic.
There are two cycles at play here and they compound in opposite directions.
The distraction cycle: every time you feel cognitive friction and reach for AI, you get a dopamine hit from the instant answer. Your brain learns that friction = reach for tool. The tool removes the discomfort. The discomfort was the training stimulus. You never leave the beginner phase.
The focus cycle: every time you push through the friction, you force the pathway to activate. It gets stronger. Problems that required conscious effort start requiring less. You enter flow states more easily. Learning compounds.
Most developers are running the distraction cycle without realizing it.
Cal Newport’s research on deep work puts the average knowledge worker at roughly 90 minutes of genuine deep work per day. For developers leaning heavily on AI assist, anecdotally that number is probably lower because the reflex to outsource friction fires before you even register that you were about to think.
The good news: this is reversible. Habits built the dependency. Habits can break it.
The bad news: the first two weeks are genuinely uncomfortable, which is exactly why most people don’t make it past them.
Why focused devs have never had a bigger edge
Here’s the thing nobody talks about.
While AI dependency is quietly degrading the independent thinking capacity of a large portion of the developer workforce, the bar for standing out has dropped dramatically.
You’re not competing with exceptional developers. You’re competing with their attention spans.
The average developer in 2025 starts more than they finish. Jumps between tools. Vibe codes something, hits a wall, asks AI for a fix, gets a fix that introduces two new problems, asks for fixes for those, ends up three layers deep in AI-generated code they don’t fully understand.
The developer who can sit with a problem, actually think through it, hold the full context, and ship something they genuinely understand that person is increasingly rare.
And rare things are valuable.
Press enter or click to view image in full size
This isn’t about being anti-AI. The best developers I know use AI aggressively. They’re just using it as a tool, not an oracle. There’s a difference between “I’ll use AI to generate the boilerplate for this auth middleware while I think through the actual security model” and “I’ll ask AI what the security model should be.”
One leverages the tool. The other outsources the thinking.
The developers who figure out that distinction who keep the thinking in-house and let AI handle the execution are going to have an enormous edge over the next few years.
Because most won’t.
What I actually do differently now
I came out of the 30 days with a different relationship to AI tools, not a rejection of them.
Here’s what changed practically:
The 30-minute rule. Any problem I’m working on, I spend at least 30 minutes on it before I open an AI tool. Not because AI isn’t helpful. Because forming my own hypothesis first means I can actually evaluate whether the AI’s answer is good. Without the hypothesis, I’m just accepting whatever sounds plausible.
AI for execution, not for thinking. I use AI heavily for boilerplate, for drafting documentation I’ve already outlined in my head, for writing tests once I’ve designed the logic. I don’t use it to figure out what the logic should be.
I read error messages again. Fully. Before I do anything else.
I write things out before I prompt. If I’m going to use AI to help with something complex, I write out my own rough answer first. Even if it’s wrong. Especially if it’s wrong. The act of articulating a position even a bad one makes the AI’s response dramatically more useful, because now I’m evaluating an answer rather than receiving one.
None of this is revolutionary. It’s just deliberately putting friction back in the right places.
The part I didn’t expect
Thirty days after the experiment ended, I still haven’t gone back to the old patterns.
Not because I’m disciplined. Because the new ones feel better. The work feels like mine again. The decisions feel considered. The code feels understood rather than assembled.
That’s the thing nobody tells you about AI dependency: you don’t notice what you’re losing while you’re losing it. The decline is smooth. Each individual shortcut feels reasonable. The compound effect is invisible until you stop and look back.
I’m not telling you to quit AI tools. I use them every day.
I’m telling you to occasionally check whether you’re using them as a lever or a crutch. Because there’s a version of this where AI makes you dramatically more capable. And there’s a version where it makes you comfortable and slow and slightly confused about why your work feels hollow.
The difference is whether you kept the thinking.
Resources worth reading
- Deep Work by Cal Newport the foundational framework for this whole conversation
- Cognitive Offloading research UC Santa Barbara the actual science behind the dependency pattern
- Andrej Karpathy on Software 3.0 the most honest framing of where AI fits in a dev workflow
- The Shallows by Nicholas Carr how digital tools reshape cognition, written before LLMs but more relevant than ever
- Paul Graham Disconnecting Distraction short, worth reading
Top comments (0)