The silence AI turned off

The "is it still worth learning to code if AI exists" discussion now restarts from scratch every two weeks. I've been chewing on this for months — first as a theoretical question, then as one that decides what you tell a junior asking "where do I start". An evidence base has appeared, so it's no longer opinions vs. opinions. And, to put it mildly, both loud extremes are wrong.

This isn't "learn or don't learn". It's about what exactly you hand to AI and what you keep for yourself. And about one thing AI genuinely changed — not in the "everything got flipped" direction, but in a much subtler one: the forced feedback loop reality used to run for you is gone.

Two extremes — both one-sided

First, let's clear the table of the two loud camps that show up in every discussion:

• Camp #1: "No need to learn languages anymore — AI will write it". Loudest among those selling bootcamps "become a programmer in 30 days with AI".
• Camp #2: "Everyone will get dumber, in a few years there will be no specialists". Loudest among experienced people slightly spooked by their own helplessness in front of how quickly juniors now look productive.

A false dichotomy. At its core, nothing fundamental has changed: as always, what's valued is conscious understanding, not dry memorization. AI only deepens the gap between those who actually get it and the rest — and does it faster than before.

What the evidence says

This isn't "it seems to me" anymore. Concrete measurements have appeared across 2025–2026:

• Nature (June 2026), "Is AI ruining our skills?" — over-reliance on AI degrades skills. Not only in programmers — in doctors too.
• Anthropic RCT on 52 engineers: an AI assistant produces ~17% lower skill formation. The subtle bit — it's not task speed that suffers, it's skill formation. Short-term you're faster, long-term you're worse.
• AI & Society: deskilling is a structural problem with no off-the-shelf fix. Not "AI is bad", but "the tool changed the learning environment".
• TechCrunch: engineers refuse to work without AI because they can't anymore. Headline: "this could come back to bite them".

The conclusion from expert K. Crowston (quoted in Nature) — it's not "AI or not", it's: consciously decide which skills to keep for yourself and which to hand off. That's the line the rest of this post sits on.

What to hand off, what to keep

Concretely — here's where the line runs:

You can't review what you can't read. So "understand without knowing the language" is an illusion; it's blind trust in AI. And that's the line between "a specialist who uses AI" and "an AI operator". The first drives AI; the second is driven by it.

Another distinction people constantly conflate: skill loss ≠ skill never formed. An experienced engineer who suddenly leans on AI stays strong — they already have a model of the world. A beginner who sits on AI 100% from day one will never build that model. These are different problems with different consequences.

How to learn the right way — and why, at the core, nothing has changed

You still need to learn — just the right way. Not dry memorization of terms or commands, but understanding the logic and having a sense of how the whole thing works. The value isn't in "remember", it's in "understand → can apply → can judge".

And the main thing: this is exactly the same as it has always been, long before AI. What was always valued were people who actually get it as specialists, not those who memorized dry. AI didn't flip anything — it just made the easy path easier (now you can "do" without understanding), so the gap between these two types only widens.

Whoever learned consciously will stay that way. Whoever was lazy will stay lazy. AI didn't create this split — it just lit it up.

Test it on simple examples

• Math: knowing the multiplication table by heart isn't yet "knowing math". The calculator killed long-column arithmetic, but not the need to understand what multiplication is and when to apply it.
• Phones: you no longer keep numbers in your head — the phone remembers them. But you do need to know who to call and why. Memory handed off — judgment kept.
• GPS: you don't keep routes in your head. But you do need to understand where you're going — and notice that GPS is taking you on some nonsense detour.

In all three the machine took the mechanical memory, not the understanding and judgment — and that's how it always was. AI is the next iteration of the same thing, just with a wider range of what you can now delegate.

Syntax is a byproduct, not the goal

This doesn't mean "no need to know syntax". You'll still pick up syntax — but as a byproduct of understanding and practice, not as the goal. You didn't learn %d with flashcards — you understood the logic of format strings, and it stuck on its own. That's how healthy learning works: understand → apply → syntax settles in by itself. AI just speeds up this loop, if you use it right.

But one thing did change: the forced feedback loop is gone

"Nothing changed" — that's about 80% true. One important thing did change, and it's not a small one.

The environment used to force you to learn. You couldn't ship working code without understanding it at least enough to make it work. The compiler, the bug, the crash — those were teachers dragging everyone to a baseline by force. The struggle was mandatory — and it built understanding even in the not-too-diligent.

Now that teacher is turned off: you can get a working result without ever understanding how it works. The immediate punishment for not understanding is gone. From that comes a consequence the "nothing changed" line misses:

The gap isn't only widening between fixed types — the new environment pulls in those who would once have learned, because the easy path now delivers a working result without immediate punishment.

So "the lazy stay lazy" is too narrow. This is more about environment and habits than character: most people react to friction and feedback, they aren't lazy "by nature". Remove the friction and even the non-lazy drift. The risk isn't only in the lazy.

And the subtlest part — the illusion of competence. AI papers over gaps so smoothly that you don't bang into the wall of your own ignorance and never notice it's there. It feels like you understand because "it all works" (the metacognitive laziness effect). So staying diligent got harder: now it requires consciously creating friction for yourself — sometimes turning AI off, trying to explain or write without it, checking "do I actually understand, or just think I do". Reality used to do this for you — now it's on you.

The rules are the same, but the game got sneakier — the easy path now disguises itself as success.

For those who already get it — the game is different

Everything above is about forming a skill from scratch. If you're already formed — the picture is flipped, and it's worth saying out loud, otherwise the post reads as "AI is bad", which isn't the point.

An experienced engineer knows their edges. They see what AI handed them → in half a second they spot where it's "off", where it's "yes, but not for this case", where it's "yes and fine here and now". They have a built-in "this doesn't smell right" detector assembled from years of real incidents. AI doesn't turn that detector off — on the contrary, it feeds it faster: more examples, more cases, more patterns to filter quickly.

So the same AI that risks turning a beginner into an operator turns a senior into a turbo-senior. Boilerplate in 30 seconds instead of 3 minutes. Exploring a new library — in an evening, not a week. Porting a module from Python to TypeScript — actually doable, not "someday later". This is nitro into an already assembled engine.

And one more subtle thing: AI can't lie to an experienced engineer unnoticed. AI lies to them just as often as to a beginner — they just catch it before they finish reading the sentence. What for a beginner is a trap of the illusion of competence, for a senior is just one more bug report in the daily stream.

For those who already put in the time, AI is turbo. For those who haven't put it in yet — it's one more reason not to. There's no middle ground between these two states — and that makes "when exactly you learned" one of the most important factors in a career.

Systemic risk: the broken ladder

Seniors come from juniors. If AI eats junior work and nobody builds the foundation → in 5–10 years there will be nobody to grow into a senior. The "lost generation" of specialists risk is real — and that's not alarmism, it's a simple piping problem: the input pipe is blocked, the output one isn't.

Why this won't go to zero (but will hurt): the "everyone stopped learning" equilibrium is unstable. Scarcity makes those who did learn overvalued → the incentive to learn comes back. Companies that consciously grow specialists get a moat. But the transition is painful: operators will fall flat at the first serious incident, when AI lies clearly and convincingly and there's nobody around to tell the truth from the lie.

My position

• Learning the language is mandatory — aimed at reading and judgment, not memorization. You can't shortcut the road to reading fluency; AI doesn't shorten it, it just quietly makes it optional.
• While the majority "hands understanding" over to AI — learning right now is a window of opportunity: you become a scarce specialist who drives AI, not one who competes with it. It's a short-term window — the market will adapt sooner or later. While it's open, it's a sin not to use it.
• The healthy mode: use AI to go faster and learn deeper, not to avoid learning. The anti-pattern is "do the whole task for me". The pattern is you write it yourself, AI explains "why", and you make sense of its variant before accepting it.
• A paradox for the road: to use AI effectively, you have to know enough that you don't need it for the simple stuff. Whoever depends on AI for simple things will be fooled by AI on hard ones with no chance of catching it.

Sources

• Nature — Is AI ruining our skills? Early results are in (June 2026): nature.com/articles/d41586-026-01947-1
• InfoQ — Anthropic study, ~17% skill mastery drop: infoq.com/news/2026/02/ai-coding-skill-formation
• TechCrunch — Coders are refusing to work without AI: techcrunch.com/2026/05/29/coders-refusing-ai
• AI & Society — AI deskilling is a structural problem: link.springer.com/article/10.1007/s00146-025-02686-z