I read a headline last month: "Developers Think AI Makes Them 24% Faster. The Data Says 19% Slower."
I closed the tab. Opened it again. Read it three times.
I've been using Copilot and Claude every day for over a year. I genuinely could not tell you where my thinking ends and the autocomplete begins anymore. That scared me more than the statistic did.
So I did something dramatic: I banned myself from all of it for five days.
No Copilot. No Claude. No ChatGPT. Nothing. Just me, the official docs, Stack Overflow, and whatever I could remember from actually learning things.
Here's what happened.
Day 1: Slower. But Weirdly Calm.
I'm working on a FastAPI backend — authentication flow, JWT refresh tokens, the usual. Normally I'd type a comment like # generate a JWT with 24h expiry and return it and watch Copilot draft the function. Instead, I opened the PyJWT docs like it was 2019.
It took me three times as long to write the same function.
But something strange happened: I wasn't anxious. Usually I'm moving fast enough that there's a low hum of "am I sure this is right?" running in the background. Today I was slow enough that I actually read what I wrote. I understood it. The hum was gone.
I finished maybe 40% of what I'd normally finish. I wasn't panicking yet.
Day 2: The Bug That Took 2 Hours and Changed Something
This is the day that mattered.
I had a token refresh endpoint that kept returning 401s even when the token was valid. Normally my workflow is: describe the bug to Claude, paste the stack trace, get three hypotheses, test them, move on. Total time: maybe 20 minutes.
Today I had none of that. Just me staring at a stack trace.
I spent two hours on it. I went through the JWT decode logic line by line. I added print statements like a person who learned to code before debuggers existed. I read the RFC on token expiry. Twice.
The bug was a timezone issue — datetime.utcnow() vs datetime.now(timezone.utc). Classic. Embarrassing.
But here's the thing: I found it. By myself. And when I found it, I felt something I genuinely hadn't felt in a long time. Not pride exactly — more like ownership. Like I could defend every line of that code in a courtroom.
I sat there for a second after fixing it. Just sat there.
Day 3: The Uncomfortable Realization
By Wednesday I had a working theory about what I'd actually been doing with AI tools.
I wasn't using them to go faster. I was using them to skip the uncomfortable part.
You know the part I mean — the 10-15 minutes where you're not sure what approach to take, where you're holding multiple incomplete thoughts at once and none of them have resolved yet. That liminal zone between "I see the problem" and "I know what to do."
I had been prompting my way out of that zone every single time.
And prompting your way out of discomfort is not the same as solving a problem. You get the answer. You don't get the understanding.
Day 3 I spent a lot of time just thinking. Sitting with open questions. It felt incredibly inefficient. It also felt like what programming used to feel like when I was getting good at it.
Day 4: I Had an Idea That Was Mine
Thursday I was working on restructuring how the API handles user sessions — whether to store state server-side or lean fully stateless with the JWT payload.
I made a call: keep the server-side session store, but slim it down to only store revocation data. Stateless for validation, stateful only for invalidation.
Not a revolutionary decision. Junior devs make this call every week.
But I could trace exactly why I made it. I'd held the tradeoffs in my head for two days. I'd thought about load patterns, refresh frequency, what we'd need to scale. Every part of the decision had a reason I could articulate without referring to a chat window.
That feeling — being able to trace your own reasoning — I hadn't had it in a while. I hadn't noticed it was missing.
Day 5: The Actual Numbers, and the Real Finding
Friday I did a rough accounting.
I shipped somewhere around 15-20% less than I normally would. Not the catastrophic 24% the headline promised. The productivity hit was real but not devastating.
But that wasn't the interesting finding.
The interesting finding was that I'd forgotten how to be comfortable not knowing things.
AI tools had quietly trained me to treat uncertainty as an input/output problem. Not sure about something? Describe it. Get output. Move on. The discomfort of not knowing becomes a prompt. The prompt becomes an answer. The answer becomes code.
Except — the discomfort of not knowing is not a bug in the programming process. It's the process. That's where the thinking lives.
I'd been skipping it so consistently that when I had to sit in it again, it felt almost unbearable. I kept reaching for a keyboard shortcut that wasn't there.
That's not a productivity problem. That's a confidence problem. I'd outsourced the hard part so many times that I was no longer sure I could do the hard part.
What I'm Actually Doing Now
I'm not quitting AI tools. That would be performative and dumb.
But I'm scheduling one AI-free day per week. Wednesdays. Not a full ban — I'll still use docs, Stack Overflow, search. Just no generative AI.
Think of it like the gym. You don't go to the gym because it's more efficient than not going. You go because the resistance is the point. The muscle doesn't build any other way.
My ability to sit with ambiguity, hold open questions, trace my own decisions — that's a muscle. It had atrophied. I hadn't noticed because I was shipping fine.
Shipping fine is not the same as thinking well.
The Challenge
Pick one day next week. Remove the AI layer entirely.
Not to prove something. Not to be productive. Just to find out what's actually still there when you have to think without the net.
You might be fine. Or you might find the same thing I found: a gap where your confidence used to be, so slowly eroded you'd stopped noticing it.
Either way, you'll know. And knowing is the point.
If you want to try this yourself, I built ai-detox — a CLI that tracks AI-free coding days, logs daily notes, and shows your streak.
pip install ai-detox
ai-detox start
ai-detox log "Spent 2hrs on that bug — found it myself"
ai-detox end
ai-detox status
Top comments (0)