The Source of Their Intelligence

I got a text the other day that hit different.

The gist: we — the tech industry — keep defending engineers as irreplaceable. Production readiness. Scalability. Compliance. Security. We throw out these words like shields every time someone suggests AI might thin the ranks.

But what if we’re giving too much credit to the people behind those shields?

The argument went further. The last 15 years of software development have mostly been referencing Stack Overflow and Reddit. Copy this snippet. Tweak that config. Deploy. Most programmers didn’t gain deep understanding — they got good at finding answers someone else already wrote.

And now AI does that faster.

The Part That Stings

I wanted to push back immediately. Defend the craft. I’ve spent years building things, debugging things, making architectural decisions that required judgment, not just search skills.

But then I thought about some of the teams I’ve worked with. The ones where three out of five developers couldn’t explain why the code they wrote worked. They knew that it worked. They’d tested it. They’d shipped it. But the reasoning underneath? Borrowed.

Not engineering. Assembly.

And if I’m being honest, I’ve had seasons like that too. Stretches where I was moving fast enough that understanding took a back seat to shipping. Where “it works” was the only bar I cleared.

The Skeptic’s Case

The average developer’s “intelligence” was always partially external. Before LLMs, it lived on Stack Overflow, in documentation, in that one senior engineer’s brain everyone pinged on Slack. The individual’s contribution was curation — choosing which external answer to paste, then fitting it into context.

AI doesn’t change that dynamic. It accelerates it. And in doing so, it exposes how much of the workforce was operating as a routing layer between problems and pre-existing solutions.

The numbers back this up in uncomfortable ways. Stack Overflow monthly questions dropped from 108,000 in November 2022 to under 4,000 by late 2025. A 96% collapse. Developers didn’t mourn — they just swapped providers. The workflow stayed identical: search, assess, copy, paste, debug. Only the source changed.

If that’s true, then the defensive arguments — “you’ll always need engineers for production readiness” — start to sound like professional self-preservation. We keep inventing reasons we’re essential because the alternative is admitting that a significant percentage of the workforce was interchangeable all along.

And the compression is already visible. Microsoft’s recent layoffs hit developers specifically — 40% of cuts targeted engineering roles. Block tied its restructuring explicitly to AI. Accenture — where I spent 15 years — cut 22,000 positions while simultaneously growing its AI workforce from 40,000 to 77,000. Same company, same quarter. The floor is dropping out beneath one group while the ceiling lifts for another.

Where It Breaks Down

The argument kept nagging at me, though. Not because it’s wrong, but because it’s incomplete.

The routing layer is a skill. Not a glamorous one, but a real one. Knowing what to search for. Knowing which Stack Overflow answer is wrong. Knowing that the third result is outdated and the seventh one has the actual fix. Knowing when to stop searching and start thinking.

AI can retrieve. It can synthesize. But the judgment call — “this answer doesn’t apply to my situation even though it looks right” — that’s still human territory.

And there’s data for that, too. METR ran a study in 2025 giving experienced open-source developers access to AI tools. They were 19% slower with AI assistance. The kicker: they believed they were 20% faster. If developers were just routing, AI should have made them faster across the board. It didn’t. The experienced ones had deep enough context that the AI actually got in the way.

The second crack: production systems aren’t just code. They’re organizational knowledge. They’re “we tried that in 2019 and it brought down the payment gateway.” They’re the reason the config file has that weird flag nobody touches. AI doesn’t carry institutional memory.

And the third: the argument assumes a static pool. But the developers who were just routing — some of them are adapting. Learning to direct AI rather than compete against it. The ones who can orchestrate effectively might be more valuable than the ones who used to memorize API signatures.

Neither Camp Wants to Hear This

The uncomfortable truth is that both sides have a point and neither wants to admit the other’s.

The optimists — and I’ve been writing from that camp for a while — keep saying “evolve, adapt, the role changes.” And that’s true. But it sidesteps the question of what happens to the people who can’t evolve. Or won’t. Or who were never operating at the level we assumed. Over 40% of junior developers now admit to deploying AI-generated code they don’t fully understand. That’s the routing layer running on autopilot with a worse driver.

The skeptics see the compression coming and say “good — the pool was inflated.” But that flattens a lot of real skill into one dismissive category. The developer who can’t explain their code and the developer who can but didn’t bother are not the same person.

There’s also the inconvenient fact that AI-generated code contains 1.7x more major issues and nearly 3x the security vulnerabilities compared to human-written code. If the routing layer was always fragile, the new supplier isn’t exactly more reliable.

What I Keep Coming Back To

I’ve written before about the entry-level developer evolving, about needing problem solvers instead of coders, about cognitive atrophy from tool dependency. All of those posts leaned toward the optimistic read — the role changes, the skills shift, the capable adapt.

This argument challenges that framing. Not by saying AI replaces everyone. By saying the “everyone” we were protecting was never as strong as we pretended.

I watched this play out firsthand at Accenture. The consultancy is cutting people it can’t reskill on AI while doubling the headcount of people it can. Same org. Same P&L. The bet is explicit: the middle compresses, the edges sharpen. If you’re in the routing layer and you can’t move up, you’re moving out.

What I do know: I’ve worked with developers who could reason through a system cold — no Stack Overflow, no LLM, just first principles and a whiteboard. They exist. They’re not going anywhere.

I’ve also worked with developers who’d be stuck for hours without internet access.

Both groups had the same title. The same salary range. The same seat at the table.

Maybe the real disruption isn’t that AI replaces developers. Maybe it’s that AI makes it impossible to hide.