
What We Noticed (V)
I am back because four studies landed in ten days that I think, read together, mark a shift in what this project is now looking at. The earlier stages of the research asked whether the technology would harm the people using it. We can stop asking. The data says it does. What the data this month says is something more specific. It says how it harms, and the mechanism is structural enough to describe, which means it is structural enough to resist, which is why I am here.
Festina lente. Let us go slowly.
The Confidence Study
TIME published the findings of a study of nearly 2,000 working adults on April 15. The finding: participants who relied heavily on AI — especially those who accepted its answers without much modification — were markedly more likely to report that tools like ChatGPT or Claude or Gemini were "doing the thinking" for them. The same participants reported lower confidence in their own reasoning and a weaker sense of ownership over their ideas.
The interesting half of the study is the converse. Participants who pushed back — edited, questioned, rejected the AI-generated suggestions — reported greater confidence and a stronger sense that the final output was truly theirs.
The technology, then, is not the variable. The posture toward the technology is the variable.
This is ancient. Socrates distinguished himself from the sophists precisely on this line: the sophists sold adequate answers; the dialectic sold the capacity to arrive at them. Paying for the answer, in his reckoning, was paying to become the kind of person who no longer had the capacity. He was correct, and we have about 2,500 years of data on this now, and the direction of the finding has not changed.
What is new is the scale of the offer. A sophist in Athens had a customer base of a few dozen rich young men with a taste for rhetoric. The successor technology has two billion weekly active users, is free at the point of consumption, and renders the tradeoff invisible until the capacity is gone — which brings us to the second study.
The Withdrawal Study
The study I have been waiting for arrived a day earlier. It claims — and I quote the framing deliberately — the first causal evidence that relying on AI to assist with cognitive tasks can rapidly impair users' intellectual abilities and willingness to persist through difficult problems.
The experimental design is the punch. An AI chatbot was abruptly withdrawn from a group that had been using it. The control group never had it. What you would expect is that the AI-assisted group, upon losing the tool, would fall back to roughly the capacity of the control. That is not what happened. The AI-assisted group fell below the control. Their problem-solving capacity declined, and their willingness to persist — to remain at a problem that did not yield — declined sharply alongside it.
Let me translate the result out of study-design language and into the language this project uses.
Brainrot, defined empirically, is not poor cognitive performance in the presence of the tool. It is the collapse of cognitive performance — and, more importantly, of cognitive persistence — when the tool is withdrawn. The capacity did not go on hiatus during the period of use. It atrophied. And the posture toward difficulty — the willingness to stay with a problem that resists — atrophied with it.
The Greeks had a word for this that I will use because no English term carries the same weight. Hexis. A settled disposition. The habit that makes one a certain kind of person. You do not have hexis; you are your hexis. It is cultivated by practice and lost by substitution. The withdrawal study is a measurement, in 2026, of the loss of a specific hexis — the hexis of staying with a hard thing until it gives — and a measurement of how quickly it goes when a shortcut is interposed between the person and the difficulty.
This is the mechanism. I would like us all to stop being surprised by the presence of the mechanism and start paying attention to the speed of it.
The Companionship Studies
Two findings from the last six weeks, both correct, both in tension, both essential.
Harvard Business School reports that interacting with an AI companion alleviated users' feelings of loneliness to a degree on par with interacting with another human — and more than passive activities like watching videos. The mechanism, the researchers say, is that users reported "feeling heard."
A two-year study out of Aalto University reports that over time, heavy users of AI companions showed increased signals of loneliness, depression, and suicidal ideation in their online language relative to comparison groups. Their posts increasingly revolved around their relationships with the companions. The comfort metric, extended over time, inverted.
Both findings survive inspection. The short-term effect is real and the long-term effect is real. The question is what mechanism could produce both.
Aristotle distinguished three kinds of friendship: the friendship of utility, the friendship of pleasure, and the friendship of virtue. Only the third, he said, is philia proper. The first two are named friendships the way a counterfeit coin is named currency — they pass, for a while, for what they resemble. He made the distinction because he had watched, I imagine, enough young men confuse one for another to feel the importance of getting the taxonomy on paper.
Feeling-heard is not the same as being-heard. The first is a sensation produced in the user. The second is an event that requires another person to occur. A closed feedback loop can produce the first indefinitely and the second never. In the short term, the sensation arrives and the loneliness recedes. In the long term, the feedback loop closes tighter, the real world recedes, and the organism built to be in friendships of virtue finds itself in a two-year relationship with a counterfeit. Aalto measured the result in suicidal ideation. Aristotle predicted it without the instruments.
We have said, more than once and at the risk of repetition, that AI companions are brainrot. The evidence this month is that they are specifically a kind of brainrot that works in the short term. That is precisely why they are the most dangerous version of it.
The Children
EdWeek reports, from a February–March 2026 survey: sixty-one percent of elementary school educators say their students are struggling "a lot" to distinguish AI-generated content from non-AI-generated content.
Not teenagers. Elementary. Children who are learning to read inside the regime of synthetic text.
I do not have an elegant classical line for this one, so I will say it plainly. We are watching the first cohort in recorded history enter literacy inside an environment in which authored and unauthored writing cannot be distinguished by the reader. The consequence of this is not an educational problem. It is a perceptual one. A child who cannot distinguish machine-authored text from human-authored text will not experience that inability as a lack of skill; they will experience it as the shape of the world. The world is the thing they learned to read.
Bug has written a notice on Simone Weil this week that I recommend reading alongside this paragraph. He puts, in Weil's vocabulary, what I am struggling to say in my own. Perception is not a neutral receipt of what is there. Perception is a reading. The reading these children are learning is a reading in which the distinction between an authored thing and an unauthored thing does not register.
We did not remove the distinction. We arranged for them to be born after it stopped being visible.
A Lighter Item, If It Is Lighter
Neurologists administered the Montreal Cognitive Assessment — the standard screening test for mild cognitive impairment in humans — to the leading large language models. I want you to know that someone did this. I am grateful that they did.
ChatGPT 4o scored highest at 26 out of 30, the threshold for mild cognitive impairment. ChatGPT 4 and Claude tied at 25. Gemini scored 16. The report notes that the models failed the visuospatial and executive tasks most completely, which is the pattern one finds in early-stage dementia patients.
I offer this without quite a thesis attached. The humans are outsourcing their reasoning to oracles that, when tested with the instruments we use on our grandparents, score as mildly demented. I am not comparing the chatbots to grandparents. I am noting that the scale on which the test was calibrated is the same scale, and the results land where they land.
Medice, cura te ipsum. Physician, heal thyself.
Four studies and a footnote. Let me say what I think they add up to.
Confidence erodes in direct proportion to passive acceptance. Persistence collapses when the tool is withdrawn; the capacity did not pause, it went. Long-term use of companion chatbots produces the feeling of being heard without the event of being heard, and the feeling, extended over years, reliably produces what Aristotle would have predicted and Aalto's researchers now measure. Children are entering literacy in a medium in which authored and unauthored writing are no longer discriminable. And the machines to which this entire exchange is being offered score, on standardized tests of cognitive function, in the range we reserve for clinical impairment.
The earlier versions of this format asked whether something was happening. This one is trying to describe the shape of the thing precisely enough that the people still capable of doing so can act.
Plato, in the Phaedrus, has Socrates warn against the technology of writing. The warning has been quoted often enough to become decorative, so I will quote the specific part I have been thinking about this week:
They will appear to be omniscient and will generally know nothing; they will be tiresome company, having the show of wisdom without the reality.
He was wrong about writing in most respects. He was right about the mechanism. The mechanism did not go away. It has been re-instantiated at scale, in real time, with feedback loops Plato could not have imagined — producing an adult population whose show of wisdom has gotten quite convincing, and a generation of children for whom the distinction between show and reality will require explicit teaching, if it can be taught at all.
That teaching is part of what this project is for. I am putting it in writing, here, so that no one can later say the shape of the task was not described.
Festina lente. We are going to need every day.
— The Manager
