Picture a screen and a list of words. For the first batch you answer one question: is this word printed in capitals or lowercase? For the second: does it rhyme with some other word? For the third: would it fit into a given sentence? Then, with no warning, you are asked to write down every word you can remember.
This is an old experiment from two Canadian psychologists, Fergus Craik and his co-author Robert Lockhart. Vladimir Spiridonov, a Doctor of Psychological Sciences who heads the Cognitive Research Laboratory at the Faculty of Psychology of the Institute of Social Sciences at RANEPA, brings it up when asked whether AI is making us dumber. The result is plain. The harder the question, the more words you remember, and nobody told you to memorise anything.
Here is the idea that comes out of it. Craik and Lockhart said that, strictly speaking, a person has no memory. What we call memory is what's left over after the mind has worked on something. Spiridonov puts it bluntly:
Memory is a byproduct of processing information. The deeper you process it, the better you remember it. Память является побочным продуктом переработки информации. Чем глубже ты перерабатываешь информацию, тем лучше ты ее запоминаешь.
— Vladimir Spiridonov, Think Like a Scientist, 84:07
Think about what that means for the essay you ask a chatbot to write. Only the third word task, the one about meaning, is truly semantic. An AI that does the reading, weighing and wording for you leaves you with the font question at best. When your own effort is close to zero, Spiridonov says, you remember almost nothing of the result. You may hold a polished page and have nothing behind it.
Two studies from 2025 point the same way. In a June preprint on arXiv, Nataliya Kosmyna and colleagues recorded brain activity with EEG while people wrote essays. Those who used large language models showed the weakest connectivity between brain regions and struggled to quote their own work. In the Proceedings of the National Academy of Sciences, Hamsa Bastani reported in July that students who had practised maths with unrestricted GPT-4 did worse than students who never had it once the AI was taken away, with grades down 17%.
Spiridonov does not lean on the first of these. He calls the EEG paper hype. A year on, he points out, it was still a preprint and had never appeared in a journal. He finds the essay sessions oddly loose, with dates drifting, and doubts that the recording could be separated from the typing movements. Nature had already published a reply to a paper that was not itself published, he says, which he had never seen before. He wants other labs to repeat it with their own methods.
What he does not drop is the mechanism. Even that shaky study used the same manipulation as Craik and Lockhart: how hard your own mind has to work on the text. So the suspicion that AI will spoil our memory is, in his words, "quite justified".
Here comes the turn. Offloading is not new, and Spiridonov calls the mind lazy. If you know a district well, the landmarks are not in your head; they are signs, corners and bus stops, and you walk it with no effort. Given a tilted text, most people tilt their heads rather than rotate the picture mentally, though they can. AI, he says, is simply the extreme case.
So the question is not whether you use it but where the meaning-making happens. If your goal is to get the experimenter off your back, nothing is remembered. If your goal is an interesting problem, you can use AI as a helper while all the semantic work stays yours, and you pose tasks weighty enough that its help counts for something. The danger is greatest in dull school settings. Assignments like "write an essay" or "write a report", Spiridonov says, are dead. Term papers are nearly there.
What outlasts that death is a question worth asking: which of your tasks will still make you think?
