Usually, during the first lesson or lecture on programming, I tell my students that they should not use AI for learning and writing code. Could this be a manifestation of neo-Luddism?
To understand the topic, I have read quite a few studies from universities and research centers regarding the effectiveness of using AI.
To put it very shortly — because of AI, it seems to you that you are getting smarter, getting information and results faster — and that is true. But over a distance of more than a couple of days, you, on the contrary, become dumber, lose most of the acquired knowledge, and additionally even worsen the cognitive abilities of your own brain and the willpower to solve complex problems.
And if your brain and willpower were so great that they withstood the onslaught of neural-slop until 2026 — I invite you to read the full article, and who knows, maybe it will change your approach to self-education.
Learned Helplessness
An MIT Media Lab study — the university divided 54 students into 3 different groups. The first group wrote essays using LLMs, the second used internet search, and the third wrote completely on their own.
The study lasted for 4 months, and after each essay, they analyzed the students' brain activity. The work of different parts of the cerebral cortex, the sense of authorship, and the ability to at least approximately repeat what was written. For 3 months, the third group practiced writing essays through trial and error, without even having access to the internet. Meanwhile, the first could use AI as an assistant and teacher to work out better quality. Right from the beginning of the experiment, it became clear that the first group strained their brains much less, and without overcoming difficulties, our brain does not form new neural connections. And 3 months later, when these groups were swapped — that is, the neuro-dependent wrote essays on their own — their results were ten times worse than those who wrote on their own all this time, or used a search engine.
To help you understand the scale of the disaster — in the AI group, 83% of participants could not remember anything about their own essay, which they had finished writing just a few minutes before the question. Whereas for the other groups, this percentage was around 10% (which is also a lot, but apparently the requirements were not exactly easy).
It is also an interesting fact that writing with the help of search turned out to be less effective than independent work.
I know that many people criticize this and similar studies. For example, they point out the small sample size. Also, the process of using AI, the model, what prompts were used, etc., is not completely clear. Also, as a counterweight, there are papers that show the opposite effect — the benefit of AI in learning. Next, we will look at a couple of them.
The Straw Benefit of AI in Education
A recent Harvard study showed the advantage of an AI tutor over live, active classroom sessions. Active sessions are not lectures, but lessons with discussion, debates, group work, and are considered one of the best pedagogical practices.
For the experiment, 184 people from the physics department were selected. They passed a test, and then for 2 weeks, the first half went to the classroom, while the second half stayed home and studied with AI. After 2 weeks, they passed the test again — as a result, the AI group showed 20% less time spent on learning and about a 20% better result.
It would seem — AI rules, and it's time for teachers to become delivery drivers. But there is nothing surprising about the fact that students preparing specifically for a test got a better result on the test than students studying the subject as a whole, holding debates and discussions, which undoubtedly took time away from cramming. To be fair, they should have taken a group that would just read a paper sheet of questions and answers all these 2 weeks; I am sure its result would have been the highest.
The case from J-PAL MIT has another problem. Schoolchildren from Brazil who had access to an AI-core learning platform showed, on average, results 9% better than schoolchildren without one. But the thing is, this platform is not just an AI. It is a full-fledged LMS with a study program + AI. Here, AI might not have played any positive role at all, but the very presence of a preparation program for the Brazilian exam naturally gave an advantage over students who did not have such a program. Many companies create such a straw man.
For example, the Embry-Riddle study (2026) showed huge progress of students from using AI. This study is actually as typical as it gets. If you read it in more detail, it turns out that the control group there did not study at all and did not perform any tasks — and only relative to it did the group studying with the help of AI win. Outran a stone — well done.
The Calculator Effect
I also often see opinions like: "Calculators were once rejected too." I want to say only one thing — they did exactly the right thing.
Russia and China still do not allow the use of calculators in mathematics both in schools and in universities — and in these countries, the level of fundamental mathematics is much stronger than in countries where calculators were allowed (which is clearly confirmed by the world PISA rankings from the OECD).
At the same time, calculators themselves are not banned: they can be used in physics, chemistry, biology, and later at work.
Those who did not reject calculators have already regretted it, and some countries are now rolling back. For example, the UK officially banned the use of calculators in primary school because children forgot how to count. In the US, the level of mathematics is also at a catastrophically low level, and Western teachers themselves admit this mistake.
I advise you to read about the school RCT experiment by Bastani et al. on the "illusion of mastery," as well as about the "lying calculator" experiment and what the "calculator effect" is, if you haven't heard of it before. There are a lot of similarities with AI here (which is also confirmed by a large-scale analysis from the RAND Corporation).
I am not calling on everyone to give up AI today. No. I am not even calling for using neural networks "rationally," as this is a meaningless phrase loved by AI companies. I offer a simple solution, proven by the calculator era: during training in a certain specialty (for example, programming), the use of AI for solving problems related to this specialty must be COMPLETELY banned.
AI is not a calculator, it covers all subjects at once, so for each individual subject, the ban is different. In biology, you cannot use AI for biology, but you can, for example, use it to find literature — I think you get the logic.
Thus, by refusing the help of AI in the part of the subject being studied, you will get preferences from neural networks in the form of a quick search for information on additional disciplines and personal matters, but you will not "liquefy" your brain in the area in which you want to develop.
Conclusion
- Using AI to solve learning tasks — be it writing an essay or code, noticeably makes you dumber and helpless against complex tasks.
- Studying with AI is better than complete inaction, and you can also use it to prepare for tests slightly more effectively than going to lectures.
- In order not to lose your brains — you need to refuse to use AI in the area you are studying.
Don't take a calculator to a math class).
P.S. I wrote this post without neural networks, because I want to learn how to write posts, and not just get the best result here and now.
Top comments (0)