The thing that every dev I know is quietly worrying about is not AI taking their job. It's losing the skill to do their job.
We're not talking about it out loud because it feels like admitting weakness, and because these tools are genuinely powerful. You can roll out a feature in an afternoon that previously took you a week. Autocomplete guesses what you're thinking. A chat agent lets you restructure that flaky module you've been avoiding for months. But then in that in-between moment, the third or fourth prompt, the third or fourth tab-tab-tab-accept cycle, that quiet question pops up: could I still do this from scratch?
Now don't get me wrong, I'm not saying that devs couldn't write a for loop if they wanted to. I'm saying that this part is what's fading. The deep work of wrestling with a problem, thinking it through, fumbling around in the dark until the solution crystallizes. Because instead of spending fifteen minutes thinking through the problem, the model gives you the answer in fifteen seconds, so what you're doing is no longer building something. You're reviewing it, which is a far less taxing and less creative task.
The research is starting to catch up with this concern. In a Microsoft study presented at CHI 2025, 319 knowledge workers showed a clear trend: the higher the trust in AI-generated content, the lower the level of critical thinking. The change in workflow is not about problem solving, but about checking and integrating. In another study, this time from MIT Media Lab about essay writing, people who used LLMs were the ones with the worst brain connectivity. They called it cognitive debt. Even productivity is more doubtful than the hype says: in a 2025 study by METR, experienced open source developers using AI tools turned out to be slower than those who didn't, but convinced themselves of the opposite.
All the devs I know have felt this, not as some abstract fear, but as a concrete experience of reaching for an answer and realizing the shelf is empty. I don't need a study to tell me that the knowledge I used to have is slipping away. That the syntax I once produced on autopilot is now a few keystrokes and a prompt away. That the kind of instinct only a developer who's been through dozens of bad architectural decisions can have is now being outsourced to a model that has seen a million architectures but never lived inside one.
The scary part is that it happens under the radar. You're still writing good code, because the model writes good code. You're still getting good performance reviews. You're still shipping. Until you're not, and that's when the tools don't work: when a weird bug shows up, when a novel constraint hits, when an edge case the training data never saw appears. Only then do you realize if there's actually a developer under the hood, or if it's just a workflow.
No one wants to go back to the old days of writing boilerplate the hard way. But the devs who are quietly preparing are doing it in the not-so-cool way: writing some code without AI, wrestling through the confusion, keeping those weak muscles working. The scary part is not that AI will replace you. It's that one day you'll wake up and you'll have replaced yourself, one accepted suggestion at a time.
