DEV Community

Cover image for They Learned to Code Before Copilot. They're Not Anti-AI. They're Pro-Evidence.
Debashish Ghosal
Debashish Ghosal

Posted on AI-assisted

They Learned to Code Before Copilot. They're Not Anti-AI. They're Pro-Evidence.

#ai #employeeexperience #career #growth

In July 2025, METR published a study that nobody in a boardroom wanted to read.

They took 16 experienced open-source developers and 246 real issues from repositories those developers had worked on for years. Each issue was randomly assigned: AI allowed, or AI not allowed. With AI allowed, the tasks took 19% longer.

That's not the part that stuck with me. This is: before starting, the developers expected AI to speed them up by 24%. After finishing, after being slowed down, they still believed it had sped them up by 20%.

Hold on to that gap between feeling and fact. It runs through this whole series.

If you started your career somewhere between 2015 and 2020, before autocomplete could write a whole function, this series is about you. Or about the engineers you manage. It isn't a memoir. It's an attempt to put real data next to the stories we tell about "the skeptical senior dev."

First, Let's Not Misuse That Study

Before anyone screenshots "AI makes developers 19% slower": METR has flagged those results as out of date.

In February 2026 they reported that their follow-up experiment had broken down, and why it broke is the interesting part. A growing share of developers refused to take part because they didn't want to work without AI, even at $50 an hour. Between 30% and 50% admitted they were holding back tasks they didn't want to do by hand. METR's own read is that developers are probably sped up more now than in early 2025, but that its data is "only very weak evidence" of how much.

One participant put it this way: going back to the old way felt like "trying to get across the city walking when all of a sudden I was more used to taking an Uber."

So in about a year, the same group went from "this helps, I think" to "please don't make me work without it," and the measuring stick broke in the middle. Keep that in mind. It's what the rest of this piece is about.

The Numbers Say Skeptical, Not Hostile

The 2025 Stack Overflow Developer Survey (49,000+ responses) splits developers by experience. The mid-career band, 5 to 10 years, is the group I care about here.

Their sentiment is positive: 62.7% of mid-career developers feel favorable about AI tools. They're not holdouts.

Their trust is a different story. 45.8% distrust the accuracy of AI output. 33.1% trust it. Only 2.8% "highly trust" it.

Go one band up, to 10+ years, and the caution gets sharper. Those developers had the highest "highly distrust" rate (20.7%) and the lowest "highly trust" rate (2.5%) of any group. Stack Overflow's own interpretation was "a widespread need for human verification for those in roles with accountability."

Use it daily, trust it rarely. That sounds like a contradiction until you remember it's exactly how you treat a new hire's first pull request.

Is that skepticism? Or is it just code review?

"Almost Right" Is the Most Expensive Kind of Wrong

The biggest single frustration in the same survey, cited by 66% of developers, is "AI solutions that are almost right, but not quite." Second place, at 45%: "debugging AI-generated code is more time-consuming."

Anyone who has been on call understands why that ranks first. Code that's plainly wrong fails in CI and you move on. Code that's almost right passes review, ships, and pages you three weeks later because of an edge case nobody wrote a test for.

A mid-career engineer is in an awkward position here. You've seen enough production incidents to know where "almost" hides. You're also still the one getting paged.

There's a quieter number in that list too: 20% of respondents said they've become less confident in their own problem-solving. That's one in five, and it's a lot more than zero.

Plot Twist: The Skeptics Ship More AI Code

This is where the easy story falls apart.

In July 2025 Fastly surveyed 791 developers. About a third of senior developers (10+ years) said more than half of the code they ship is AI-generated. Among juniors (0-2 years) it was 13%. That's roughly 2.5x.

Seniors were also more likely to say AI makes them "a lot faster" (26% vs 13%), and 59% said it helps them ship faster overall, against 49% of juniors.

The cost showed up in the same data. Just under 30% of seniors said they edit AI output enough to cancel out most of the time they saved, compared with 17% of juniors.

Fastly's explanation fits what you'd expect: experienced engineers are better at spotting when code "looks right" but isn't, and that lets them lean on the tool harder.

So the most distrustful group is also the group putting the most AI code into production. That's not a paradox. They use it more because they trust themselves to catch it.

Three Kinds of Pre-AI Engineer

What follows is a framework I'm proposing. The surveys don't use these categories. But the data points in these directions.

The Auditor. Uses AI all day and reviews every line. Treats output as a draft from a fast but unreliable colleague. This is probably most of the "use daily, trust rarely" crowd, and probably most of Fastly's seniors shipping AI code.

The Holdout. Hasn't bought in, at least not for the newer agentic tools. In Stack Overflow's data, 38% of developers have no plans to use AI agents at all, and 14% use AI only in autocomplete mode. The reasons aren't only philosophical: 87% are worried about agent accuracy, and 81% about security and privacy.

The Convert. Started out skeptical, found the workflow that fit, and now won't go back. METR's broken experiment is basically a population of converts.

The question to ask about someone isn't "are they pro-AI or anti-AI?" It's "what would it take to move them one box over?" For an Auditor, maybe better tests. For a Holdout, maybe a security answer. For a Convert, maybe someone reminding them to check the output.

Why 5 to 10 Years Is the Interesting Band

This part is my theory, so read it as argument, not data.

Engineers with 5 to 10 years of experience learned the craft the old way: stack traces, bad docs, Stack Overflow threads from 2013, a senior who circled the one line in your PR that would have taken prod down. They've owned incidents. Most have been through at least one rewrite of something they wrote.

What they usually don't have is a seat in the meeting where the AI budget gets decided.

Which means they're the ones who'll maintain whatever the AI era produces. GitClear analysed 211 million changed lines of code and found that in 2024, copy-pasted code exceeded "moved" (refactored) code for the first time in the history of their data, with a sharp rise in duplicated blocks. Someone will own that duplication in 2028. It's probably not the VP who approved the licenses.

When a mid-career engineer pushes back, a lot of the time they aren't defending the old way of working. They're defending their future on-call rotation.

What the Skepticism Is Actually Protecting

Y Combinator's partners, who are hardly AI pessimists, said it plainly in March 2025. Discussing the finding that a quarter of their W25 batch had codebases about 95% AI-generated, general partner Diana Hu said that even builders who rely heavily on AI need to be good at one thing: reading the code and finding bugs.

That skill isn't typing speed. It's judgment, and judgment comes from being wrong in production a few dozen times.

Stack Overflow asked developers when, in a future with advanced AI, they'd still go to a human. The top answer, at 75%, was "when I don't trust AI's answers."

That's the job description the pre-AI engineer already has. The only question is whether their organization sees it that way.

The Takeaway

The loudest AI skeptic on your team might be your best AI user. Look at their merge history before you judge them by their Slack takes.

So tell me:

  • Which of the three are you right now: Auditor, Holdout, or Convert?
  • Has your trust in AI output gone up or down in the past year, and what changed it?
  • If you're 5-10 years in, does "use daily, trust rarely" match your week, or is that already outdated?

Next in the series: your CEO sees 5x productivity, and your engineers see a longer review queue. Both might be right.

Sources

Top comments (1)

Some comments may only be visible to logged-in visitors. Sign in to view all comments.