It's 2 a.m. Production is down. You're staring at a function you "wrote" three weeks ago, except you didn't write it — you described it, an AI produced it, the tests passed, and you moved on. Now there's a bug in the middle of it, and you realize, with a cold drop in your stomach, that you have no idea how this code actually works. The feature shipped on time. The understanding never arrived. This is the bill coming due, and it has a name.
A new kind of debt, with a familiar interest rate
In March 2026, Addy Osmani, an engineering leader at Google, named the thing many of us had been feeling but couldn't quite point at: comprehension debt. It's a new category of technical debt, specific to the age of AI-generated code — not the usual debt of messy architecture or skipped refactors, but the hidden cost to your own understanding when you let a machine do the thinking and skip the part where you do it yourself. Like any debt, it's invisible at the moment of borrowing. The code runs. The demo lands. The interest accrues quietly, and it comes due at the worst possible hour.
Same speed, less understanding — the receipt
If that sounds like a sermon from someone who codes by candlelight, the data says otherwise. In a randomized controlled trial, 52 software engineers were asked to learn a new library. The AI-assisted group finished in roughly the same time as the control group — no heroic speedup on the clock. But on a follow-up comprehension quiz, they scored 17 percentage points lower: 50% versus 67%. And here's the part that should make your palms sweat: the steepest decline wasn't in some abstract conceptual zone. It was in debugging — the exact skill you reach for at 2 a.m. You bought speed you didn't even get, and paid for it in the one ability you most need when things break.
Liu Weipeng saw this coming, just not in code
Long before AI could autocomplete a whole module, a Chinese programmer named Liu Weipeng wrote a quiet, stubborn book called Dark Time. His central idea: real learning doesn't happen in the seconds it takes to find an answer. It happens in "dark time" — the invisible hours your mind spends chewing, deriving, turning a problem over. The productive part of learning is background processing you can't see, and it's exactly the part AI offers to skip for you.
Liu has a rule that reads like a prophecy for the AI era: think before you search. When you wrestle with a problem first, you dig a hole in your mind — a question-shaped cavity. Then, when the answer arrives, it drops in and sticks, because something was waiting to receive it. Grab the answer instantly and there's no hole, nothing to catch it; it slides right off. "Immediately searching for the answer," he warns, "is making you dumber." Swap "searching" for "prompting" and he could have published it last week.
Generation is not comprehension
Here is the distinction the whole problem hinges on, and it's the same one Liu drew between producing and understanding. Being able to make code appear — by typing it, or by asking a model to type it — is not the same as being able to explain it. Liu's test for true understanding was brutally simple: teaching is the best learning. If you can't explain a thing clearly, you almost certainly don't understand it. An AI can hand you a flawless, finished block, and you can ship it, and you still won't be able to teach it to the junior next to you — which means, by Liu's measure, you don't actually know it.
And it's not just whether you use AI — it's how. The same trial split developers by behavior, and the gap is enormous. Those who used AI as pure generation delegation — generate, accept, move on — scored below 40% on comprehension. Those who used it for conceptual inquiry — asking why, weighing tradeoffs, probing the reasoning — scored above 65%. Same tool, opposite outcomes. The figure below lays the two paths side by side.
Google's Addy Osmani named it 'comprehension debt' (March 2026): a randomized trial of 52 engineers learning a new library found AI-assisted coders finished just as fast but scored 17% lower on a comprehension quiz (50% vs 67%), with the biggest drop in debugging. The split is in HOW you use the tool — generation delegation scores under 40%, conceptual inquiry over 65%. Liu Weipeng's Dark Time names the cause: real learning happens in the thinking time you just skipped.
What this means for you tomorrow morning
So should you stop using AI? No — and that would be the wrong lesson. The fix isn't abstinence; it's using AI the way Liu describes learning. Make it answer questions instead of just producing artifacts: ask why it chose this data structure, what the tradeoffs are, what breaks if the input doubles. Then close the chat and explain the code back, out loud or in a comment, in your own words — if you can't, you've found your comprehension gap before prod does. And at least once, debug a piece of AI code by hand, no model in the loop, until you've felt how it actually behaves. Remember Liu's other warning: you remember what you've retrieved, not what you've re-read. AI agents now generate code 5 to 7 times faster than a human can comprehend it, which means the bottleneck — and the debt — is no longer typing. It's understanding. Don't let the swipe outrun the comprehension. Weave the new knowledge into the web you already have, or it isn't yours.
You can delegate the typing. You can't delegate the knowing.
Generation is the swipe; comprehension is the bill — and only your own dark time pays it down.
Framework drawn from Liu Weipeng's Dark Time (刘未鹏《暗时间》). "Comprehension debt" named by Addy Osmani (Google) in an O'Reilly / Medium piece, March 2026. Trial figures: a randomized controlled trial of 52 software engineers (50% vs 67% comprehension); generation-delegation use scoring below 40% vs conceptual-inquiry use above 65%; AI agents generating code 5–7× faster than humans comprehend it. Popular tech commentary.
2026 年 3 月,谷歌的工程负责人 Addy Osmani 给一种很多人隐隐感到、却指不出来的东西起了名字——理解债(comprehension debt)。这是 AI 生成代码时代特有的一类技术债:它不是架构混乱、不是欠下的重构,而是当你让机器替你思考、跳过本该自己走的那一步时,欠下的对"自己是否真懂"的隐性成本。和一切债务一样,它在借的那一刻看不见。代码能跑,演示漂亮,利息却在悄悄滚动,最后挑你最扛不住的时辰来收。
速度没快多少,理解却少了一截——这是收据
如果这听起来像个点蜡烛写代码的老古板的说教,数据会替你纠正。一项随机对照试验里,52 名工程师被要求学一个新库。用 AI 辅助的那组,用时和对照组基本一样——秒表上并没有什么神勇的提速。但在随后的理解测验里,他们低了 17 个百分点:50% 对 67%。而真正让人手心冒汗的是:跌得最狠的不是什么抽象的概念区,而是调试——正是你凌晨两点最仰仗的那项能力。你买来的速度其实没拿到,代价却恰恰是出事时你最需要的那项本事。
刘未鹏早就预见了,只是没说是代码
早在 AI 能自动补全整个模块之前,一位叫刘未鹏的中国程序员写了一本安静而执拗的书——《暗时间》。他的核心想法是:真正的学习不发生在你找到答案的那几秒,而发生在"暗时间"里——你的大脑反复咀嚼、推导、把一个问题翻来覆去想的那些看不见的时辰。学习中真正高产的部分,是你看不见的后台运算,而那恰恰是 AI 主动提出要替你跳过的那一段。
刘未鹏有一条像是为 AI 时代写的预言:先思考,再搜索。当你先和问题死磕,你会在脑中挖出一个洞——一个问题形状的空腔。等答案到来时,它正好落进去、卡得牢,因为早有东西在等着接它。若你一上来就抓答案,脑里没有洞、没有接口,它就直接滑走了。"一遇到问题就立刻搜答案,"他警告,"是在让你变笨。"把"搜索"换成"提示词(prompt)",这话像是他上周才发的。
那么该停用 AI 吗?不——把它读成"戒掉"就学错了。解药不是禁欲,而是按刘未鹏描述学习的方式去用 AI。让它回答问题,而不只是产出零件:问它为什么选这个数据结构、有哪些取舍、如果输入翻倍会先崩在哪。然后关掉对话框,用自己的话把代码讲回来——出声讲,或写进注释——如果讲不出,你就在生产环境之前先抓到了自己的理解缺口。而且至少有一次,要亲手调一段 AI 代码,全程不让模型插手,直到你真切感受到它到底怎么动。别忘了刘未鹏另一句告诫:你记住的是你检索过的,不是你重读过的。如今 AI 生成代码的速度,是人类理解速度的 5 到 7 倍,这意味着瓶颈——以及那笔债——早已不是打字,而是理解。别让"刷卡"跑赢"看懂"。把新知识织进你已有的那张网里,否则它就不是你的。