Table of contents (9 sections)
“You did nothing, the AI did it.” People say this to me almost every time I show them a page of this site. I am a web developer and I run web-developpeur.com. I publish reading notes (a book summarized idea by idea), a course on the Go programming language and articles. I work with Claude, an AI from Anthropic that writes text and code.
The AI types a lot of the code and text on this site, and I don't hide it. But I am fed up with this sentence, because it confuses typing with doing. It mostly comes from people who know nothing about the subject and who couldn't redo the work, even with the same AI. It forgets what makes a page right, clear and useful: someone read it, judged it, refused it, corrected it.
The repository of this site is the folder where the git tool stores every version of the code. It holds 2,382 commits, records of changes, all under my name. None of them says whether I typed the line or the AI wrote it, and none keeps a trace of what I refused. That repository is private: nobody but me can open it, and even I couldn't read this work in it.
The calculator stood trial too, the piano never did
In 1986, the Christian Science Monitor, an American newspaper, covered the great calculator debate among American math teachers. Their association, the National Council of Teachers of Mathematics (NCTM), had been recommending calculators in class since at least 1980. Fewer than half of the teachers use one, and often only to check an answer. For a first-grade teacher, the calculator is a crutch: “The risk is, once you have a crutch, you rely on it more and more.”
That crutch objection was serious, and it still holds. Here is an example I built: you type 48 × 520 instead of 48 × 52. You read 24,960 instead of 2,496. The calculator did nothing wrong, you mistyped. To see it, you need to know the result should be around 2,500. That reflex of estimating is the part of math the machine does not do for you.
So doing math with a calculator is doing math differently, as long as you keep that reflex. The comparison holds for AI too, with one caveat. A calculator always computes exactly what you typed. An AI can answer wrong with as much confidence as right. Your share does not shrink: with an AI, you watch what you ask, and also what it answers.
Nobody dares put the piano on trial. You don't tell a pianist: “You did nothing, the piano plays the music.” Everyone knows what the piano does not do. It doesn't pick the piece, it doesn't interpret anything, and it can't tell when a note sounds wrong. Neither can an AI. One difference remains: a piano plays nothing without a pianist, while an AI produces something even when nobody can play. That is where the crutch objection finds its target.
When the AI is wrong and nobody can tell
Each time I correct the AI, I write the rule in a corrections file, stored in the repository, so it doesn't repeat the mistake. That file is part of its skills: rule files it rereads before working. The table below comes from it. Its last two rows come from my reading note on Computer Networks: A Systems Approach. This networking textbook by Larry Peterson and Bruce Davie (Systems Approach LLC, 2020) is free online. A reading note is split into ideas: sections that each carry one notion from the book.
| What the AI had done | What I said | What changed |
|---|---|---|
| It had applied my “smart child” skill (write for someone who doesn't know the subject) only to the passage I had pointed out. | “Did you use the smart child skill?” (it answered: no) | The skill is rerun on the whole reading note, and an automatic count flags sentences that are too long. |
| It had added an example presented as mine: resuming a job at step 6 of 11. | I judged it useless, and I have never done that. | An example “from my own experience” must come from real use. An invented scenario, with invented numbers, gets removed. |
| It used the word “link” (the connection between two machines on the network) fifteen times in the reading note, without ever defining it. | “What exactly is a link?” | Before a reading note is handed over, we list the repeated common nouns and check that a passage defines each one at its first appearance. |
| It had piled three patterns, five acronyms and a story into one idea of the reading note. | “I understood nothing of chapter 8 of the book. If a child who knows nothing reads it, I'm not sure they'd come out with anything.” | That idea is rewritten around a single concrete action, opening a web page, unrolled in three steps. |
None of these errors is a typo. Every sentence was readable and plausible. The flaw was elsewhere: a hollow, invented or overly dense sentence, and only a reader who knows what the page must deliver can see it. In the second row, the AI had lent me an experience I never had. Without review, that is a fake with my name on it.
My loop: five steps, only one for the AI
Here is the loop I follow every time, and how it splits the work between the AI and me. Judging means reading and asking whether it is true, useful and clear. Checking means stepping out of the text: redoing the calculation, reopening the source.
Two stories: method matters, feeling proves nothing
Two stories show where the result is decided when a machine helps: one about method, the other about the feeling of having worked well.
The first comes from chess. In 2005, the site Playchess.com ran a “freestyle” tournament: everyone played with whatever chess programs they wanted, alone or in teams. Groups of grandmasters, the best players, signed up, each with several computers. The winner was a pair of American amateurs running three computers at once. According to Kasparov, their strength was knowing how to steer their computers to look very deep into the positions. Garry Kasparov told this tournament in 2010 in The New York Review of Books, and drew a formula from it. A weak player, a machine and a better method beat a strong computer alone. They also beat a strong human with a machine and a worse method.
The second comes from the stopwatch. In 2025, the research organization METR had 16 experienced developers work on 246 tasks, in projects they had known for five years on average. For each task, a draw decided whether AI was allowed. Before starting, the developers expected AI to save them 24% of their time. Afterwards, they estimated they had saved 20%. Measured: their tasks took 19% longer, so they lost time. The authors specify that the result covers the tools of early 2025, not every case.
I don't conclude that AI always slows you down: the study covers specific tools at a given moment. I conclude that the feeling of going fast proves nothing, so you have to measure. In both stories, what decides the result is not the tool but how you use it: the method at Kasparov's tournament, the measurement at METR. In another article, I measured what my way of working with agents costs.
Five questions to tell whether someone did the work
Here is the test I ask myself, and that you can ask anyone who works with an AI:
- Can they explain every choice without rereading their exchanges with the AI?
- What did they refuse? An AI always proposes, so work with no refusal at all is suspect.
- What did they try to break (pushing it to its limits), redo by hand, prove wrong?
- Who answers if it is wrong? A person, under their own name, not the machine.
- What rule did they add so the error does not come back?
There is also a test to give whoever tells me “the AI did it”: hand them the same AI and the same subject, then look at what they hand back. If they copy the first answer and publish it, they fail all five questions, and for them the sentence is accurate. I can answer each one with a precise case: the corrections in the table above are my refusals and my rules.
What the AI really does
It writes first drafts, small checking programs (scripts) and a large share of the lines on this site. It reads a whole book much faster than I do. I decide what needs checking, and I look at what the checks answer. I would gain nothing by hiding this split.
This article is written with Claude too. The theses and the arguments start from what I told it. Claude looked up the sources and wrote the text, and I reread and approve before publication. Facts that come from elsewhere have their source at the bottom of the page: you can check them without taking my word for it.
My take
“The AI did it” closes the discussion instead of opening it. The right question is “how did you check it?”. I answer that one gladly, with proof. The sentence “the AI did it” calls for no answer: it only says that the speaker did not look.
I think the people who throw it can't say what is wrong in an AI's answer. They don't know what to ask it, or when to say no. The machine is the same for everyone, and that is exactly why the result is not.
I will keep working with AI. I will also keep refusing what it proposes when it is wrong, hollow or invented, and nobody will see those refusals in the history.
Conclusion
Checking work has a flaw: when it succeeds, it leaves no trace in the result. A page without errors doesn't look like it took work, and that is why people believe it was easy.
The calculator, the piano, chess, the METR stopwatch: each time, the machine does not replace judgment, it moves it to where it is least visible. That is why I write my corrections in a file of the repository: it is the only lasting trace of that work.
Sources
Documentation consulted on November 15, 2026:
- The great calculator debate, Christian Science Monitor, May 9, 1986: the dispute among math teachers, the NCTM and the “crutch”.
- The Chess Master and the Computer, Garry Kasparov, The New York Review of Books, 2010: the 2005 tournament and the formula.
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, arXiv, July 2025: the controlled trial with 16 developers.
- 324 million tokens re-read for 3 million written, my article on what an agent session costs.