Coding Assistants Are Making Software Engineers Worse — And Nobody In Management Wants to Say It

Hot take: we are in the middle of a hidden skill-atrophy crisis in software engineering, and the entire industry is pretending not to smell it because slowing down AI adoption is a career-limiting move in 2026.

Let me be clear: I use coding assistants daily. I am not a Luddite. But I have watched enough junior and mid-level engineers over the last eighteen months to know that something is rotting in the foundation of our craft, and the metrics everyone is celebrating are measuring the wrong thing.

The Hidden Skill Crisis

Every senior engineer I trust reports the same pattern. Junior engineers are shipping faster than ever. Tickets are closing. PRs are landing. Velocity dashboards look fantastic. And the underlying engineering judgment — the thing that determines whether a system survives its first year in production — is collapsing.

I sat with a mid-level engineer last month who could not walk me through the auth flow of the feature they had shipped the week before. They used an AI assistant to scaffold it, it passed tests, they merged it. When I asked them to explain the token lifecycle, they froze. They had never actually read the code. They had never had to. The AI wrote it, the AI understood it, and the AI will not be on call at 2 AM when a refresh-token bug takes down the billing pipeline.

This is not isolated. Microsoft Research, METR, and Stanford HAI have published data showing developers using AI assistants complete tasks faster on paper but score measurably worse on comprehension, debugging, and architectural reasoning. The productivity numbers leadership celebrates measure typing speed. They do not measure thinking.

The GPS Effect, But Worse

We have seen this movie. GPS navigation did not make people better drivers. It produced a generation that cannot read a map. Coding assistants are doing the same to engineering, except more aggressively. GPS degrades slowly and you can still see the road. AI assistants scaffold entire abstractions — repository patterns, dependency injection containers, async state machines, ORM query plans — that the user has never had to reason about from first principles. You do not just lose the ability to read a map. You never learn that maps exist.

And the part that makes me angry: we know this is happening. We see it in code review and in the panic when autocomplete goes down for a day. And we are doing nothing to design rollouts that preserve deep reasoning, because the quarterly all-hands wants velocity, not judgment.

Yes, Senior Engineers Use Them Well. That Is The Point.

I know the counterargument. Experienced engineers use AI assistants as a force multiplier, not a crutch. They prompt, review, reject, refine. Fair.

But that judgment was earned over a decade of not having these tools. Senior engineers in 2026 know when the AI is hallucinating a method that does not exist, because they remember writing every line and learning what real APIs feel like. They have the scar tissue.

A 22-year-old joining in 2026 does not have that scar tissue and will not develop it on a substrate where the first instinct for every problem is ask the AI. That instinct, baked in early, does not produce real senior engineers in five years. It produces senior engineers in title only — people who can ship features but cannot debug their own code without a chatbot whispering hints.

The Conclusion Nobody Will Say Out Loud

AI coding tools are not net-positive for the long-term capability of the engineering profession. They are net-positive for velocity, output volume, and headcount leverage. Those are not the same things. Velocity without judgment is how you get the next CrowdStrike outage.

If you are an engineering leader and you are not designing your AI rollout to preserve deep reasoning — required cold-coding exercises, AI-free reviews on critical paths, deliberate architecture-from-scratch sessions — you are not increasing your team's productivity. You are mortgaging it. The bill will come due, and unlike a 2 AM outage, this one will not be fixable in a single sprint.

Mr. Technology writes about the parts of AI the marketing decks leave out.