Jellyfish surveyed 812 senior engineers and found that code churn — lines written then deleted within two weeks — went up 39% on AI-assisted teams. Seventy-five percent of tech leads expect medium-to-heavy technical debt within twelve months. Thirty-four percent of developers say their career path is now unclear. Productivity went up. Something else went with it. That something is the cognitive cost of reading what the machine wrote.
30-second takeaways
Writing code is encoding; reviewing AI code is decoding — and decoding always costs more, information-theoretically speaking.
Code churn grew 39% on AI-assisted teams. Seventy-five percent of tech leads expect significant debt within a year.
The burnout isn't from longer hours. It's from the specific mental work of verifying output you didn't build.
The fix isn't less AI. It's an honest count of where the cognitive work actually went.
1NEW NORMAL
After AI arrived, a developer's day quietly changed shape
The tools shipped the promise. The calendar kept the receipts.
Before AI coding assistants, an engineer's day had a rhythm you could recognize. Read the ticket. Sketch the approach. Type it out line by line, run the tests, argue with a colleague about the PR. Writing was most of the time, and the writing was also thinking. The code in your head and the code in the file were working the same problem at the same time. You knew why every line existed because you put it there.
That rhythm broke. In 2026, on an AI-heavy team, the session opens differently: type a description into the chat, wait thirty seconds, receive three hundred lines. Generation is now nearly instant. What expanded to fill the recovered time is not rest. It is review. Not the quick scan you do of your own work — where you know what every variable does and why it's there. The slower kind. Someone else's logic, written fast, in a style you didn't pick, for a problem the machine only half-understood from your prompt. Evil Martians documented exactly this shift in 2026: the job description for senior engineers at AI-heavy shops quietly rewrote itself. The headline number — ten-times productivity — stayed in the pitch decks. The footnote — "and now you review ten times as much unfamiliar code per day" — did not.
The bottleneck moved from generation to verification. The tooling, the process, and the headcount never got the memo.
2INFO THEORY
Reviewing AI code is harder than writing code — not a feeling, a mathematical fact
Wu Jun's The Beauty of Mathematics gives us the words for what's actually happening here.
Shannon's information theory draws a line between two things. Encoding: you already understand a message, and you write it out in a symbol system. Decoding: you receive a string of symbols and reconstruct meaning you didn't already have. Encoding starts with comprehension. Decoding ends with it — maybe. These are not symmetric operations. To decode something new, you have to independently reconstruct the full context. You can't just check the symbols. You have to rebuild the world those symbols refer to.
Writing code is encoding. You have the problem model in your head. You translate it, step by step, into the symbol system the compiler reads. Each line costs roughly one unit of externalizing something you already understand. Reviewing AI-generated code is decoding. The AI had a problem model — assembled imperfectly from your prompt — and translated it into symbols. To check whether those symbols are right, you can't just read them. You have to build an independent problem model and compare. That reconstruction is not cheaper than the original encoding. In most cases it's more expensive, because you're doing it defensively — against output that may contain plausible-looking errors you have no memory of putting there.
Here's the thing. Reading and verifying a 100-page math proof takes about as long as devising it in the first place, sometimes longer. The proof is sitting right there on the page, complete, but to trust it you have to re-derive it in your head. Same asymmetry, AI-generated code. The machine wrote fast. The person checking cannot take shortcuts the machine never needed.
Writing your own codeEncoding — you already hold the problem model
Reviewing AI codeDecoding — must reconstruct the model independently
Cognitive cost ratioDecoding ≥ Encoding in almost every information-theoretic analysis
3TECH DEBT
When 39% of the code is "AI churn," it stacks into one large future invoice
Speed on the front end doesn't disappear. It defers.
Code churn is a proxy for code written before the problem was fully understood. Developer writes a function, realizes it's wrong, deletes it within two weeks — that's churn. Every project has it, at some baseline rate. Jellyfish found AI-assisted teams were churning 39% above baseline. The most plausible explanation isn't that AI writes bad code in an obvious way. It's that AI writes code that looks correct, passes the quick scan, gets merged, and then surfaces its flaw two sprints later when it runs into a piece of context the model didn't have. By then, the original code is woven into four downstream modules. Unwinding it is expensive.
Jellyfish Developer Productivity Report 2026 · Code churn grew 39% with AI-assisted teams; 75% of tech leads expect medium-to-heavy technical debt within 12 months; 34% of developers see no clear career path—and 32% are considering leaving the field entirely.
The 75% debt prediction from tech leads is the other shoe. They're not predicting bad code exactly — they're predicting code nobody fully understands, accumulating faster than any team can refactor. The debt isn't in the syntax. It's in the missing mental model. Every function merged without a human who built the understanding from scratch is a future incident waiting for the right context collision to set it off. At 39% higher churn rates, you're not just shipping more features. You're minting more future surprises per sprint than your incident-response team was ever sized for.
4WHAT THIS MEANS
The answer isn't less AI — it's an honest accounting of where the real work went
A hammer that doubles your nail rate also doubles the inspection time. You'd plan for that. Plan for this.
The diagnosis here is not "AI bad, go back to typing everything yourself." The speed is real. It matters. What happened is that the speed got measured and celebrated, while the cognitive transfer — from building to verifying — went largely unpriced. That gap is now showing up as burnout numbers, churn rates, and tech lead anxiety about debt. The useful response isn't a moral position on AI. It's a process question: what does it actually cost to absorb AI-generated code into a codebase you can reason about, and are you putting that cost anywhere in the budget?
Three adjustments make a real difference immediately. First, treat review time as load-bearing work, not overhead. The review at the end of the sprint, speed-run before the Friday deploy, is exactly where the 75% debt prediction comes from. If AI generates in minutes what used to take hours, those hours didn't go away. They moved to the review column. That column needs real time blocked for it. Second, keep at least one human who built the mental model from scratch on every significant piece of AI-generated code. That person is not a bottleneck. They are the link between the code in the repo and an actual understanding in a human brain. Without that link, you own code nobody can reason about when something breaks at 2 a.m. Third, track churn and treat rising churn as an early-warning signal, not just a velocity number. Churn that outpaces your review capacity is technical debt accruing at exactly the pace it takes to become unmanageable. The 39% figure is an average. Some teams are already past the point where catch-up is affordable.
The real ask is not harder work or fewer tools. It's honest work: count the cognitive cost of what you're doing, budget it deliberately, and stop calling verification overhead when it's actually the job. Wu Jun's point in The Beauty of Mathematics is that information has a cost arithmetic can describe but intuition frequently undercounts. In the AI-coding context, the undercounted item is not the tokens generated. It is the independent semantic reconstruction every reviewer must perform on every function they didn't write. That cost is, by information theory, at least as large as the original generation, and in practice often larger. You don't fix a wobbly table by sanding the top.
Ten-x code generation at thirty-percent review coverage is not a productivity gain. It's a scheduled debt payment wearing a costume.
Price the decoding. It doesn't go away because no one put it in the sprint plan.
取材:吴军《数学之美》(Shannon 信息论部分);Evil Martians《AI-assisted engineers are burning out》(2026);Jellyfish 开发者生产力报告(2026);数据以原报告为准。
代码 churn,说白了就是"问题还没整明白就写出来的代码"的代理指标。你写了个函数,发现写错了,两周内删掉,这就叫 churn。哪个项目都有这个,有个基准率。Jellyfish 发现 AI 辅助团队的 churn 比基准高了 39%。最合理的解释不是 AI 明摆着写出了烂代码——而是 AI 写的代码看起来没问题,过了快速扫视,合并进去了,然后等两个 sprint 后,碰到一块模型当初没有的上下文,问题才冒出来。那时候那段代码已经钩进了四个下游模块,拆起来很贵。
这里的结论不是"AI 不好,回去自己一行行敲"。速度是真的,而且很重要。问题是:速度被量了、被表扬了,而那个认知工作的转移——从自己建东西变成防御性地验别人的输出——没人核算过。这才是倦怠数字、churn 率、技术负责人焦虑的根。有用的回应不是站道德立场说 AI 好不好。是个流程问题:把 AI 写的代码消化进你能讲清楚的代码库,这实际上要花多少钱,你算进去了吗?
你别看这事像是在加班或者少用工具,其实都不是。就一件事:诚实。把认知成本算清楚,有意识地排进预算,别再把验证叫"开销"——它就是这份工作本身。吴军在《数学之美》里说的是,信息有个成本,算术能描述它,直觉老是低估它。在 AI 写代码这事上,被低估的不是生成了多少 token,是每个审查者对每一个他没写的函数要独立重建语义——这个成本,信息论上至少和原始生成一样大,实践里往往更大。图啥?省了打字的时间,多出来一倍验证的账单,这买卖你得算清楚。
30% 审查覆盖率下实现十倍代码生成——那不叫生产力提升,那叫伪装成效率的分期付款。
解码的成本算进去。没人写进 sprint 计划,不代表这笔账消失了。
取材:吴军《数学之美》(Shannon 信息论部分);Evil Martians《AI-assisted engineers are burning out》(2026);Jellyfish 开发者生产力报告(2026);数据以原报告为准。