I read my own commits like a stranger | Egor Fedorov<br>When MIT researchers put EEG caps on people writing essays with ChatGPT, 83% of the AI group could not quote a single sentence from the essay they had just finished. The number stayed abstract to me until I opened my own repo and did not recognize the code on screen.

What happened with me?

Yesterday I opened my project repo on github to fix one small bug. The bug was not critical, it was just wrongly set weights for the ranking system. I knew the architecture of my own code perfectly, where each file is and how it works. But that did not stop me from getting stuck for a minute, maybe two, reading the code of a file that I wrote with my own hands, but at the same time it felt like somebody else’s commit.

At first I thought maybe it really wasn’t me who wrote it, but these 28 lines of code, I was exactly the author. egorthinks "ranking system upd weight", last week. It was written well. If I was a team lead for myself, I would for sure praise the elegance and the simplicity of the implementation. But. I did not remember that it was me who wrote this code.

I scrolled through this code and read it like, probably, authors of original books read their own translations. The architecture and the idea of the implementation are mine, but damn, I did not recognize the lines of this code.

Maybe you will say that I’m just a dumb vibecoder or something like that. I thought about myself the same way, because before AI this did not happen. I really thought the problem was in me, because nobody else from my friends and colleagues talked about it.

Of course, as it turned out later, nobody said anything because, just like me, they thought the problem was in themselves. I found this out after a small survey of seven of my colleagues (4 confirmed, 2 more yes than no, and 1 denied it, but he really writes everything by hand and hates AI). It does not count as a scientific paper, but it lets me make a conclusion that the problem exists. And anyway, the scientific papers already exist.

What the MIT study found

While researching this problem I came across a paper from June 2025 in which a team led by Natalia Kosmyna from MIT Media Lab put EEG helmets on 54 people and asked them to write an essay. They split them into three groups: the first group wrote without help, the second with Google, the third used ChatGPT. One specific number caught me in this work, 83%. 83% of people from the third group could not reproduce a single quote from the essay they wrote a few minutes ago . From the same 17% who remembered at least something, they could not reproduce the text word for word. Meanwhile in the groups without AI the failure rate was only 11%.

The electroencephalography data showed the same picture as the oral survey of the participants. The brain of the no-AI group showed the strongest and most branched neural networks while writing the essay. The Google group was in the middle. And the ChatGPT group had the weakest connectivity, especially in the alpha and beta ranges, which are especially connected to internal attention in the human brain.

The paper is still a preprint on arxiv, it (as of 27.05.2026) has not passed peer review yet. There is also criticism of the sample and of the methodology of the EEG research. But even with all the criticism of the paper and the problems with EEG research, you can take the results seriously just from the oral surveys alone. When almost everyone (83 percent) of the people using AI could not give an answer about their own text, this is exactly the problem that me and my friends saw on ourselves with coding.

So why is simply accepting changes not the same as writing them by hand?

I think this happens on the lowest motor level. Our whole life we learned to write, on paper, on screen, no difference. Our thought always went into matter through our hands perfectly. When you write code you already remember the next lines of code in advance, because they were already there and you just reinforced this memory by pressing the keyboard keys, confirming the correctness of the action in your thoughts. You can ask “but what about talking?“. And I can ask back “how many conversations from today do you remember compared to messages in the messenger?“. I can’t answer for you, but for myself I can answer for sure. Conversations I remember 50/50, but what I wrote I remember almost 9 out of 10.

Meanwhile when writing code with AI, whether it’s simple autocomplete or an agent like claude code, the motor part of the subconscious accepting and saving the code in the head does not work. It’s a simple yes/no accept loop like when updating packages (does anyone even read what the terminal programs ask us to give our consent to?), only with your code.

What I described above (you remember information stronger when you create something, than when you only recognize it) is called the generation effect. Slamecka and Graf showed it on word pairs in 1978. It was...