AI was supposed to replace developers. Here's what the research actually shows.

The headlines didn't match the follow-up

For most of 2024 the narrative was uniform. McKinsey, Goldman Sachs and the big AI vendors were all projecting dramatic automation of software work. Layoffs followed. Entire engineering functions were restructured around AI-first pipelines.

Two years in, the research coming out of actual developer studies tells a more useful story. Not one of replacement, but of rebalancing. Here's what we've read, and what we've seen ourselves running a small agency that uses AI on every project.

AI-generated code ships with measurably more bugs

CodeRabbit's December 2025 "State of AI vs Human Code Generation" report analysed 470 open-source pull requests. AI-authored PRs contained 1.7 times more issues than human-only PRs (10.83 per PR vs 6.45). Logic and correctness issues rose 75%, security issues 1.5 to 2x, readability issues more than 3x.

In parallel, GitClear's 2025 code-quality study looked at 153 million lines of code and found code churn, the share of lines reverted or rewritten within two weeks, roughly doubled from the pre-AI 2021 baseline. Code duplication went from 8.3% of changed lines in 2021 to 12.3% in 2024, roughly 4x growth.

The useful takeaway isn't "AI bad, humans good". It's that the output of an AI coding tool needs review, correction and often rewriting. That work has a real cost, and it is almost always absorbed by experienced engineers.

The productivity paradox is real, and also more nuanced than the headlines suggested

In July 2025, METR published a study that measured 16 experienced open-source developers working on large repos (22k+ stars, 1M+ lines of code). When allowed to use AI tools, they were on average 19% slower than when working without them. The developers themselves expected a 24% speed-up, and even after experiencing the slowdown they still believed AI had sped them up by 20%.

The mismatch between perception and measured reality was, for many of us, the most honest finding in the study.

It's worth noting that METR updated its results in February 2026. A newer cohort with a revised study design showed a smaller slowdown, around 4%, with a wider confidence interval, and their updated conclusion was that "AI likely provides productivity benefits in early 2026." The tools improved. The workflows around them improved.

Both findings coexist. AI coding tools can slow an experienced developer working in a large, unfamiliar codebase, and can speed up the same developer working in a familiar one on a well-scoped task. The difference is supervision, context and task fit, not the tool itself.

The industry is reshaping roles, especially at the junior end

The labor side of the story is where the pendulum is most visible.

A 2025 LeadDev survey found that 54% of engineering leaders plan to hire fewer junior developers because AI copilots let seniors cover more ground. Employment for software developers aged 22 to 25 dropped nearly 20% from its late-2022 peak by July 2025. Entry-level postings are down roughly 67% from 2022.

At the same time, Gartner's February 2026 report predicted that 50% of companies that cut customer service staff due to AI will rehire for similar functions by 2027. That finding is specifically about customer service, not developers, but it captures the same underlying pattern: over-delegating to AI and then walking it back when the costs surface.

So the dev market in 2026 is shifting senior, not shrinking. Teams with strong senior engineers paired with AI are pulling ahead. Teams that assumed AI could substitute for the people they just laid off are rebuilding, usually at a premium.

What we're seeing at Revolter

We use AI tools every day, on every project we ship. Scaffolding components, drafting tests, translating content between Swedish and English, generating images, writing the first draft of a migration. It's how we deliver at the pace of a much larger team.

But a few things have become clear from the inside:

• Review is work. When we let AI write a non-trivial function, reading and verifying it takes a meaningful fraction of the time writing it ourselves would have. The win comes from the tasks where review time is small compared to the baseline, not from assuming AI output is free.
• Context is the bottleneck. AI tools are most helpful when we've already made the architectural decisions. They're least helpful when we haven't. The judgment call about what to build, and what shape it should take, still sits with the human.
• Juniors are not obsolete, but their first year looks different. Some of the boilerplate tasks a junior used to do in their first six months really have been automated. The craft they need to learn has shifted toward prompting, evaluating, reviewing and integrating, not away from it. Teams that skip hiring juniors entirely end up with a broken pipeline three years out. That's a problem worth thinking about at the industry level, not a reason to not hire.

What it means if you're weighing AI in your own team

• If you're budgeting an AI-assisted project, budget review and integration time, not just generation time. The tool is cheap. Turning its output into something production-ready is not.
• If you're hiring, a senior engineer using AI tools effectively outperforms the same number of juniors using the same tools. That's the shift the market is pricing in, and it matches what we see in practice.
• If you're tempted to cut an experienced engineer because you've seen AI ship impressive demos, keep in mind that the people with context on your systems are the ones who can actually benefit from AI the most. They're also the ones you can't cheaply replace.

AI did not replace developers. It moved the ratio of craft to judgment upward. Teams that plan around that honestly are the ones doing well in 2026.

If you want a team that's candid about where AI genuinely helps and where it quietly adds to the bill, get in touch.