Quick map: this is the long-form essay. There’s a tighter Medium version, a 9-minute video, and a 60-second cut. Read first. The AI-Resistance Self-Assessment link is at the end.
On outsourcing grief, the seven-second window, and what I learned the hard way when I stopped
I want to write about something I have been avoiding writing about for a while, because I am still inside the embarrassment of it. Long-time readers know I don’t usually publish from inside the embarrassment. I wait. I let the thing settle. I write the version where I’ve already understood what happened. This one I’m writing while I am still figuring it out, because if I wait until I have understood it, the part that is useful to anyone else will already have evaporated.
So. Here it is.
The kitchen
I asked ChatGPT what to text my sister after our mother’s diagnosis.
I want you to picture it accurately, because the accuracy matters. It was a Tuesday in October. About 9:40 in the evening. The kitchen was dark because I had not bothered to turn the light on when I came in. The kettle was on, even though I hadn’t decided to make tea — the kettle is a reflex, the way other people’s reflex is to open the fridge. My phone was on the counter. I had been holding it for the entire walk back from the car.
The diagnosis had come earlier that afternoon. I had been told by my father over a phone call that lasted four minutes. I had not yet told my sister, because she was at work, because she works in a job where you cannot be told this kind of news in the middle of the day and continue functioning, and I had decided — I thought I had decided — that I would wait until she was home and then call her.
But I knew I was going to text first. Some short thing. Are you home, can you talk. That kind of message.
And I picked up the phone, and I opened ChatGPT instead of the messages app, and I typed something close to: help me draft a text to my sister telling her mum has [diagnosis], I want to sound calm and not freak her out before we can talk on the phone.
I hit enter before I noticed I had done it.
It gave me three options. Warm. Concise. “Validating,” which is a word it uses a lot. I copied option two and pasted it into the messages app and sent it.
Then I stood in the dark kitchen and the kettle clicked twice and I felt something I could not name for about ten minutes. When I could name it, the name was wrong.
I felt relief.
Relief that I had not had to find the sentence myself. Relief that I had not had to sit with the diagnosis for the seven seconds it would have taken to write the first word of it down on my own. And then, almost immediately afterwards, I felt sick that the relief had come first. Before the grief. Before the sister. Before anything that should have been there instead.
“I felt relief — and then I felt sick that the relief had come first.”
I want to say something here, before I go on, because I know how this sounds. I am not telling you this story because I think I am uniquely terrible. I am telling you this story because I suspect a lot of you have done some version of it, and that the version you did was also small, also seemingly harmless, also impossible to fully articulate the wrongness of in the moment. It is the impossibility of articulating the wrongness in the moment that this essay is about.
The two weeks I kept doing it
The first thing that happened, after the kitchen, was that I did not stop.
I want to make this clear, because I think the convenient version of this story would be one where I had the kitchen moment, recognized something profound, and walked into a fourteen-day fast from the model. That is not what happened. What happened is that for the next two weeks I kept doing it, in smaller and smaller ways, with diminishing self-awareness, and the only thing that changed was that I started noticing afterwards.
I drafted a toast for a friend’s wedding by handing the model three details about him and letting it write the rest. The toast was fine. People laughed. One of the laughs was at a line I did not understand was a joke until I heard everyone laughing.
I asked the model to “soften” a text I was sending to my partner during a fight, because I wanted to be the bigger person, by which I meant I did not want to do the work of becoming the bigger person. I wanted to seem like the bigger person. The model is excellent at this. It is a kind of competence that does not require the underlying state.
I asked it to write a birthday message to my dad. My dad is sixty-eight. I have about, on a generous estimate, twenty more birthdays to write him birthday messages for. I outsourced one of them to a model that has never met him.
These are the ones I remember. I am sure there were more I do not.
The therapist’s office
Three weeks after the kitchen I told my therapist about it. Her name doesn’t matter for this — I’ll call her by what my notebook calls her, which is, I’m a little embarrassed to tell you, just “T.”
T’s office is on the second floor of a building above a sandwich shop. The room smells, very faintly, of the sandwich shop. I have been seeing her for about a year and a half. She is, by some distance, the best money I have ever spent.
I was telling her the story about the diagnosis. I was mid-sentence when I checked my phone, and I did not notice I had checked it. She did. She let me finish the thought. And then she said:
“You’re describing the message you sent. I’m asking what you would have written.”
I went quiet. She did not lean forward. She did not pause meaningfully. She is not that kind of therapist. She just waited until I noticed there was no answer.
I sat there for what felt like a very long time. I tried to come up with the sentence I would have written, if I had written one. I could not do it. Not because the feeling wasn’t there. Because the feeling did not yet have language. And the model, in the kitchen, had given me language before I had developed any of my own — language that fit the situation well enough that the situation had closed around it, like a wound healing over a piece of grit.
She waited me out. I said something stupid. She let me say it. Then she said the other thing she said in that session, the one I almost didn’t write down because it sounded too simple:
“Whatever you would have written, it would have been yours. The message you sent is no one’s.”
I want to give you a number, because the numbers are how this essay pays for itself. She charges $250 an hour. That sentence cost me about four dollars at her rate, and a man dead for one thousand eight hundred and forty-six years had been saying a version of it, in private, for free, every night, for nineteen years of his life.
I’ll get to him.
The borrowed brain
There is a window between having a thought and outsourcing it. It is about seven seconds long. Sometimes longer. Often, for me, much shorter.
In that window, the thought is still yours. You have not paid it out yet. You could still sit with it. You could still let it develop into something that is going to surprise you, because you have not yet shaped it to fit anybody else’s mouth.
I started missing that window in three seconds. Sometimes in one. By the end of the two weeks I am describing, I was missing it before I knew I had had a thought at all.
I call this the borrowed brain. The seven-second handover. The reflex to send a thought out of your skull before you have felt the weight of it, in exchange for a tidy paragraph that does not belong to you and never will.
The borrowed brain is not a tool problem. It is the slow privatization of the part of you that used to think. (I’ve written more about this elsewhere.)
I am going to use this phrase a lot in the rest of this essay, and probably for the next year of my writing. You may as well get used to it.
The thing we are actually doing
We spent a decade arguing about note-taking apps. Whether Roam was better than Notion. Whether the second brain was really a brain or just a database with delusions of grandeur. The argument had stakes for a small population of knowledge workers who were already going to be fine, and it consumed most of the bandwidth available for thinking about how these tools were changing us.
The “second brain” debate was a category error, but a small one. It was about storage. About where to keep your ideas once you had had them.
The bigger error — the one that is happening right now, the one I am trying to write about while I am still inside it — is what we are using these models for, which is feeling.
I want to be specific, because specificity is the only way to make this argument stick.
We are drafting condolence notes in ChatGPT.
We are workshopping apology texts before sending them to people we love.
We are generating wedding speeches for our oldest friends.
We are asking the model to make this sound less angry before texting a partner we are, in fact, angry at.
We are writing eulogies. We are writing birthday messages to parents who will not be around for many more birthdays. We are drafting the texts to our siblings about our mothers’ diagnoses, and we are letting the relief that follows feel like efficiency.
I am going to make a moral claim now, and I want to be clear that I am making it as a moral claim. Every prompt is a small abdication. Most of the abdications do not matter. When you ask the model to summarize a meeting, the abdication is trivial — you would have done the same thing with a pen and a notepad if you had the time. But when the prompt is for an apology, a condolence, a goodbye, the abdication has moral weight. You are not abdicating effort. You are abdicating the thing the message was supposed to prove you had done.
“Every prompt is a small abdication. When the prompt is for an apology, a condolence, or a goodbye, the abdication has moral weight.”
I will defend this claim in a minute. First I want to tell you about a Texan.
What writing does to the body
James Pennebaker is a social psychologist at the University of Texas at Austin. (If you want to understand how to apply his research practically, this piece on journaling correctly is worth your time.) In 1986 he published a study, with his graduate student Sandra Beall, that has been replicated and refined for almost four decades since. The study was simple. He took a group of college students, divided them in half, and asked one half to write for fifteen minutes a day, four days running, about the worst thing that had ever happened to them. The other half wrote about superficial topics — what they had for lunch, what they planned to do that weekend. Then he tracked their health for the next six months.
The students in the trauma-writing group went to the campus health center significantly less often than the controls. Their immune function, when measured by lymphocyte response, showed measurable improvement. They reported, when surveyed later, that the writing had been one of the most important things they had done in college, even though many of them had wept while doing it.
Pennebaker has spent the rest of his career trying to figure out why this works. The answer, refined over the next thirty-something years, is roughly this: the act of writing about an emotional event by hand requires the nervous system to process the event on the way to the page. The writing is not the point. The processing is the point. The page is just the place the processing has to happen in order to count as having happened.
“The writing is not the point. The processing is the point. The page is just the place the processing has to happen in order to count as having happened.”
There are pieces of this that are contested, and I want to be honest about that. The size of the effect varies across studies. The mechanism is not fully understood. Some of the early effect sizes have not held up at the magnitude they were first reported. This is normal for social science and I would not want you to think I am citing it as if it were physics.
But the core finding — that writing about emotional events by hand produces measurable physical changes that typing does not appear to produce in the same way, and that talking to someone about the same events produces something different again — that finding has held up. The mechanism is the body. The body is the substrate. The page is the place the body goes to do its work.
When you prompt instead of write, you skip the processing.
The feeling does not go away. It just stays unmetabolized inside you, looking for somewhere else to come out — usually in your sleep, usually in the relationships that have nothing to do with the original feeling, usually as a low-grade something that takes you years to identify and decades to undo.
Aurelius could have asked anyone
There is a man I want to introduce, because the architecture of this essay does not work without him.
Marcus Aurelius was the most powerful person in the known world for the last nineteen years of his life. He was the Roman emperor from 161 CE to his death in 180. He had access, by virtue of his office, to every living Stoic teacher of the second century. He could, in principle, have summoned any of them to his bedchamber, at any hour, and put any question to them, and received an answer.
He chose, instead, to ask the page.
Every night, by lamp, he wrote private notes to himself in Koine Greek. They were not for publication. They were not for posterity. He did not give them a title. He certainly did not intend for me, eighteen and a half centuries later, to be quoting from them in an essay about ChatGPT. They survived because someone — we are not sure who — preserved them after his death, and because the manuscript happened to be in the right places at the right times for a thousand years.
The French philosopher Pierre Hadot spent his career on these notes. In his 1992 book La citadelle intérieure — translated into English as The Inner Citadel — he made the case that we have been reading Aurelius wrong for centuries. We have been reading him as a book of wisdom. He is not a book of wisdom. He is a record of practice. Specifically, the practice of hypomnemata — written spiritual exercises, performed nightly, intended to do something to the person doing them.
Aurelius was not writing to communicate. He was writing to process. The page was the only listener that would make him do the work. He could have outsourced it to any of a dozen brilliant living humans. He chose the friction.
I am going to say something now that I would not say in a shorter piece, because it sounds too neat, but it is true. Aurelius was doing what T charges me $250 an hour to watch me try to do. He was just doing it alone, with ink, for free, every night, for nineteen years.
There is no version of his practice that an AI could have done for him. There is no version of Meditations that could have been generated. Not because the model could not produce sentences in the right shape — it absolutely could — but because the sentences were not the point. The processing was the point. The page made him do the processing. A prompt would have done the work instead of him, and the work was the entire object of the exercise.
He could have asked anyone in the empire. He chose the page because the page is the only listener that makes you do the work.
He was not writing to communicate. He was writing to process. The page was the only listener that would make him do the work.
Renting your thoughts versus owning them
When you prompt a model, the thought is processed outside your nervous system and handed back to you as text. You become a quality-control inspector for your own inner life. You scan the output. You accept it or regenerate it. You do not carry it. The nervous system is bypassed.
When you write by hand, the nervous system is the processor. The thought passes through the body on its way to the page, and the body changes shape from having carried it. There is research, separate from Pennebaker’s, on the difference between handwriting and typing — Mueller and Oppenheimer’s 2014 paper, the famous one about laptop note-taking in lectures — and part of a broader argument I’ve made about why automation is costing us reflection, found that students who handwrote their notes retained more conceptual content than students who typed verbatim. The mechanism, they argued, was that handwriting forced reformulation. The hand cannot keep up with the lecturer, so the brain has to compress and re-encode in real time. The compression is the learning.
The same mechanism, I think, is what makes handwriting work for emotional processing. You cannot transcribe a feeling. The hand will not let you. The hand makes you compress, and the compression is where the feeling acquires shape. You go in with a vague dread and you come out with a sentence, and the sentence is yours in a way that no generated sentence will ever be, because you made it by carrying the dread through your arm.
This is the part of the argument I have been afraid to make in public, because it sounds mystical. I do not think it is mystical. I think it is biomechanical. But it sounds mystical, and I have been worried that saying it out loud will make me sound like a man who has gone slightly insane during a sixty-day phone detox.
I have not gone slightly insane. I have just stopped renting something I used to own.
You can rent intelligence. You cannot rent the part of you that grows from having had the thought.
The objection, taken seriously
I want to give the strongest version of the counter-argument now, because I have been making my own case for several thousand words and you deserve the other side.
The strongest objection runs like this. People who are not good with words have always asked other people to help them say difficult things. There is a long tradition of this. Husbands have dictated to wives. Mourners have asked priests for the right phrase. Illiterate parents have asked literate children to read and write their letters. The model is just the latest version of an old practice — the externalization of language to someone better at language. And to deny people this externalization is, the argument goes, to gatekeep emotional expression for the articulate, which is a kind of cruelty.
I take this seriously. I have thought about it a lot.
Here is where I land, and you can disagree with me in the comments and I will not delete your comment.
The difference between asking a priest for help and asking a model is that the priest is a person who knows you, or at minimum is a person who knows people like you, and who is doing something to themselves by helping you find the words. The wife who wrote her husband’s letters knew the husband. The work of writing those letters changed her, even when it did not change him. The relationship was the channel. The externalization was real, but it was externalization to another human nervous system, which was doing its own processing while it helped you do yours.
The model is not doing any of this. The model is a function. It maps your prompt to a plausible output and it does not change as a result of having done so. There is no second nervous system. There is only your nervous system, which you have just opted out of using.
This is not a small difference. It is the only difference that matters.
The line that will make people angry
There is no such thing as an AI-assisted apology. There is only an apology you didn’t make.
The same is true of condolences. The same is true of any message whose entire purpose is to prove that you went somewhere inside yourself and found something difficult to say. If a model wrote the difficult part, the message is a forgery. A useful forgery. Sometimes a kind one. But a forgery. And the person on the other end is being lied to about whether you went anywhere at all.
I know the objection. It helped me say something I couldn’t have said otherwise.
You could have said it. You chose not to learn how.
I am going to leave that there.
The notebook, what was actually in it
By day 41 I had filled 80 pages of a notebook. I went back and read it. I want to tell you what was in it, because the convenient version of this essay would be one where I just gesture at the notebook and you imagine, generously, that it was full of insight.
It was not full of insight.
About half of it was useless logistics. Grocery lists masquerading as thought. A long, boring section about whether I should change car insurance providers. A list of every meal I had eaten that week, which I had written without realizing I was doing it, and which I could not for the life of me explain in retrospect.
About a quarter of it was actively embarrassing. I am not going to describe it here. You can imagine. Anyone who has kept a journal under conditions of mild emotional pressure can imagine. There was a lot of score-settling against people who had not actually done anything to me. There was a lot of self-pity. There were two or three pages about an argument I had not had yet but had decided I was going to have, scripted out in advance, that I had then completely forgotten about by the time the actual occasion came up and which, I now realize, would have ended the friendship if I had said any of it.
But about a sixth was better than anything ChatGPT had ever produced for me. Not because it was smarter. Because it was mine, and I had not known, until I wrote it down, that the difference would matter as much as it did.
The useful pages were the ones I had been afraid to write. Every single one.
This is the part of the essay where I would normally try to draw a tidy lesson. I am not going to, because the tidy lesson is what the model would do. Instead I will tell you the one thing I learned that I trust:
A model will never embarrass you. That is the problem, not the feature. (This connects to a longer argument I made in The Case for Ordinary Proof — about why the small, specific, embarrassing detail is exactly where good writing lives.)
Two admissions before I give you the practices
The first one I have already made. I outsourced the first sentence of my sister’s grief.
The second one I have been holding back, because it is the worse of the two. About six weeks into the sixty days — I am going to keep this vague, because some details are not mine to share — I had to write something for a person who was dying, and I caught myself, again, opening the model. Not for the first message. The first one I wrote by hand. For the second one. The reply. I had received their answer, and I sat down to write back, and my hand went to the model out of pure reflex.
I did not, this time. I stopped. I wrote the reply by hand. It took me, I am embarrassed to say, about forty minutes to write four sentences.
But I want you to know that the reflex was still there at six weeks in. The model is the most addictive thing I have ever used, and I say that as someone who has used some other things. I’ve written about substance subtraction before — what happens when you remove something you’ve been using to regulate your state. The seven-second window does not magically widen because you have decided it should. It widens because you have repeatedly stood inside it without acting, and let it stretch.
Halfway in: if this is landing harder than you expected, that’s the assessment talking. There’s a short audit at the bottom — the one 312 people asked me for. Or watch the 9-minute video version first.
Three practices, in order of friction
I am going to give you three things to do. I want you to start with the one that scares you. Not the one that sounds easiest. The one you read and feel a small spike of resistance to. That is the one your borrowed brain is already starting to defend.
One. Delete the app from your phone. Keep it on desktop only.
This is the easiest, and most people will start here, and I think that is fine. The borrowed brain happens in the seven seconds between thought and thumb. Take away the thumb and roughly two-thirds of the reflex dies within a week. You will be astonished how much of your usage was thumb-driven rather than thought-driven. You will reach for your phone, and the app will not be there, and you will discover, almost every time, that you did not actually need to ask anything. The question was just the shape of the reflex.
Two. Write the first draft of any emotional message by hand. Then transcribe. (This practice sits inside a larger argument about what self-optimisation culture does to authentic human expression.)
Condolences. Apologies. Hard texts to family. Declarations. Anything where the purpose of the message is to prove that you went somewhere inside yourself. The handwriting is not precious. I want to be clear about that. You do not need a nice notebook. You do not need a nice pen. You need friction, and friction is what makes the nervous system carry the thought.
It will feel absurd for the first three messages. By the fifth it will feel like the only honest way to do it. By the tenth you will start to notice that the messages are arriving differently at the other end. People will respond to them differently. They will not be able to say why. You do not have to tell them why.
Three. A 9 PM page. One side of A5, every night, no rereading. (The tension between private writing and the pressure to perform authenticity publicly is something I explore in The Hidden Cost of Authenticity.)
Aurelius did this for nineteen years. Pennebaker’s research suggests fifteen minutes is enough to register the measurable changes. One side of A5 takes me about that long.
You do not need to know what you are going to write. The point is not to know in advance. The point is to find out by writing.
Do not reread. This is important. The temptation, after a week of this, is to flip back through the pages and look for patterns. Do not. The point is the writing, not the archive. The notebook is the place the processing happens. It is not the place the processing is stored.
Friction is not the enemy of thought. Friction is the substrate in which thought becomes yours.
What the model is actually good for
I have spent five thousand words attacking one specific use of these tools, and I owe you, before I close, a clearer statement of what I think they are actually for.
I use ChatGPT now, again, after the sixty days. I use it for things that are not me. Summarizing a long technical document I do not need to internalize. Translating between formats — markdown to HTML, prose to bullet points, that kind of thing. Writing the kind of email where the content is fully determined by context and my job is just to produce the right phrasing — scheduling, logistics, the email that is going to say what every email of that type says. Coding help. Research starting points, never endings.
These are all uses where the output is the point and the processing is not. The processing has nothing to do with me. The model is doing labor that no human nervous system was ever going to grow from doing. Take it. Use it. I do.
The thing I do not do anymore — the thing I will never do again, knowingly — is use the model in the seven-second window before I have had a thought. I do not use it to write things whose entire purpose was to prove that I had been somewhere and felt something. I do not use it for apologies. I do not use it for condolences. I do not use it for love.
This is not a moral panic about AI. It is a specific objection to a specific abdication, and I want to be precise about which one, because the imprecise version of this argument is everywhere right now and it is not helping anyone.
The kitchen, again
Months later, my sister called crying. The diagnosis had moved. There was new news, which was worse than the old news.
I was in the kitchen when the phone rang. Same kitchen. Lights on, this time.
I did not reach for the phone keyboard. I picked up the call. But before I started talking — and this is the part I keep going back to in my head, because I am not sure I would have done it if I had not been keeping the notebook — I opened the notebook on the kitchen counter and I wrote down, in shorthand, almost illegibly, the first sentence of what I wanted to say to her. Then I said it. Then I kept talking, with no other support, for the better part of an hour.
The feeling I had had that first night, the one I could not name, the one I called relief and then called wrong — I have a name for it now. It is not a comfortable name.
Grief that had nowhere to go because I outsourced its first sentence.
I am not going back. Not because the model is bad. The model is fine, for the things the model is for. I am not going back because I want to be the one who shows up.
I want to be the one who shows up. Not the one who sent something on my behalf.
If this was the version of this essay you wanted
You probably know by now that I write one essay a week. They go out on Sunday mornings, to people who wanted the longer, harder version of these arguments without the algorithm in the middle. This post is on the long end of what I publish to the blog — most of the weekly essays are tighter, more specific, more willing to be wrong on smaller questions.
There is no welcome sequence. There is no funnel. There is no upsell. There is, on Sunday, the next essay.
If you have something you want me to write about — a version of the borrowed brain you have caught yourself in, a counter-argument you think I have not taken seriously enough, a practice that has worked better than mine — the comments are open and I read all of them. I do not respond to all of them. But I read all of them, by hand, on a real screen, with no model in the loop.
That part I am keeping.
If you want the rest of the work
Three places, in order of friction:
- The Medium version — same argument, half the length.
- The AI-Resistance Self-Assessment — a twenty-minute audit that names which roles you’ve quietly handed over. Built for the 312 people who asked me for one. One PDF. No upsell.