A couple of years ago I wrote a post about the dangers of leaning on AI too hard. At the time it felt a bit like shouting at a cloud. The tools were handy, sure, but most people I knew still wrote the bulk of their own code and reached for ChatGPT when they got stuck.
That world’s gone. Pichai stood up at Cloud Next this year and said 75% of new code at Google is now AI-generated. Two years ago that number was 25%. The engineers there aren’t really writing code anymore so much as approving it. And Google isn’t a special case here. Everyone else is a year or two behind them on the same road.
Which brings me to the thing I keep noticing, in myself and in people I work with. We’re not getting better at this job. A lot of us are getting worse, and we’re doing it while shipping more than ever, which is what makes it so hard to see.
I’ve started calling it the great developer reset. Not because anyone’s starting over on purpose, but because skills that took years to build are draining out the bottom while we’re distracted by how fast everything feels. There’s a less polite name going around for the same thing. Brainrot.
Here’s the part that stopped me cold. Anthropic ran a proper randomised controlled trial earlier this year, 52 engineers learning a new Python library. Half used AI, half didn’t. The AI group finished a whole two minutes faster, which is nothing, and then scored 17% lower on the comprehension test afterwards. Fifty percent versus sixty-seven. The biggest gap of all was on debugging questions, which is exactly the skill you need when the AI hands you something that’s almost right but not quite.
Two minutes faster, a third worse at understanding what you just built. That’s not a trade I’d take on purpose, but it’s the one a lot of us are making fifty times a day without noticing.
Now, I want to be careful here, because the lazy version of this argument is “AI bad, real programmers suffer” and that’s not what I think at all. The same study found something more interesting. The people who used AI to ask conceptual questions, to understand the why, scored 65% and up. The people who just delegated the whole task and pasted the result scored below 40%. Same tool, opposite results, and the only thing that changed was how they used it.
Which means the damage isn’t baked into the tool. Lean on it to skip the thinking and you rot. Use it to interrogate the thinking and you come out sharper. Same subscription, two completely different developers at the end of the year.
This is just use it or lose it, the oldest adage in the book, except now it’s happening to an entire profession at once. I forgot most of the phone numbers I used to know by heart the moment my phone started remembering them for me. My regex used to be sharp. These days I describe what I want and let the machine write it, and if you put me in a room with no internet I’d be slower than I was in 2019. That’s not a hypothetical, I’ve tested it, and it stung.
The bit that genuinely worries me isn’t people like me though. I built my mental models the hard way before the tools showed up, so when I lean on AI I’m offloading something I already know how to do. I can feel when the answer’s wrong because the knowledge is still in there somewhere, gathering dust.
Juniors coming up right now don’t have that. There’s research out of a few different groups this past year pointing at the same uncomfortable thing. The younger cohort, the 17 to 25 crowd, leans on AI the hardest and scores the lowest on critical thinking, while the over-46s lean on it least and score highest. The explanation is the scary part. Older developers offload tasks they already mastered. Younger ones offload tasks they never learned in the first place. The neural wiring for reading a stack trace, for holding a system in your head, for that gut feeling that something’s off, it never gets built, because the struggle that builds it got skipped.
You can’t lose a skill you never had. You just become a permanent beginner with really fast output. That’s a worse place to be than rusty.
And the wild thing is the wider profession already senses this. The Stack Overflow survey from late last year had 84% of developers using AI tools, but trust in the output hit an all-time low. Only 3% said they highly trust what it gives them. Two thirds named the same top frustration, the answer that’s almost right but not quite, the one that eats your afternoon. So we don’t even trust the thing, and we’re handing it more of our thinking every month anyway. That’s the part that feels a bit unhinged when you say it out loud.
There’s a money angle to this that doesn’t get talked about enough either. The frontier models everyone actually wants to use are expensive. GPT-5.5 runs around $30 per million output tokens, and Claude’s Opus tier sits in the same postcode. That sounds like nothing until you’ve got a whole team firing off agents all day, every day, and the monthly bill lands on someone in finance who starts asking pointed questions about it.
There are cheaper options, sure. The small and open-source models have got genuinely good, and for a lot of the grunt work they’re more than fine. But that’s not what developers reach for when the problem is actually hard. When you’re properly stuck you want the smartest thing money can rent, because the frontier models are the ones pushing what’s possible in code right now, and that’s where the real intelligence gap still shows. Nobody fires up the budget model to untangle the gnarly bug at 4pm on a Friday.
So we’ve gone and built a daily habit on top of the most expensive tools going, at a point where even the big labs are reportedly losing money on every inference call. I wouldn’t bet the house on frontier tokens staying this cheap and this unlimited. The day they get metered or the price climbs, the people who kept their own skills sharp will be fine, and the ones who outsourced their whole brain to a model they can no longer afford are caught out twice over.
I don’t have a neat five-point framework for you, because I don’t think this is a framework problem. It’s a discipline problem, and discipline is boring and personal. For me it’s come down to a few stubborn habits. I read the code the AI gives me properly, not just the diff, the actual logic, before it goes anywhere near a branch. If I can’t explain why it works, I treat that as a red flag rather than a win. And every now and then I’ll build something small with the tools shut off, purely to find out which muscles have gone soft, because I’d rather find out on a Sunday side project than in an incident at 2am.
None of that makes me faster. It’s meant to stop me getting hollow. There’s a real difference between a developer who uses AI to move quicker and one who’s slowly forgotten how to do the job underneath it, and from the outside, on a good day, the two look identical. The bill for the difference just shows up later.
Use the tools. They’re extraordinary and they’re not going anywhere. Just keep doing enough of the hard part yourself that you’d still be dangerous with the power off. Most days I manage it. Some days I catch myself pasting in a for loop I could’ve written in my sleep, and I close the tab and write it myself, feeling slightly ridiculous. That small bit of friction might be the only thing standing between sharp and cooked.