My Dad died in Sept 2024 and I was the one to write his obituary. For many days, I could not produce a single word. I just couldn't get started. I really had no idea what to write. My kids and niece all suggested that I ought to use ChatGPT. Feed it the details of his life and let it do the work.
Perhaps that would have yielded something better, but I simply waited until the words, finally, came forth from my own fingers. Anything else would have felt as if I hadn't honored the memory and life of my father. I think that was an important choice for me personally.
A smart chap once said “For the things we have to learn before we can do them, we learn by doing them.” How many of my cognitive endeavors am I willing to outsource?
Thanks for the great article.
And one addendum: Cory Doctorow's enshittification of Internet platforms will leave no stone unturned:
Thanks for reading and commenting Mark, and sorry for your loss. I had a somewhat analogous experience in 2023 when my older daughter was in early infancy and obsessed with fairytales (she still is!). I asked ChatGPT to make up a custom fairytale for her. After reading the first line, I felt icky about it, turned off my phone, and made up a story of my own.
It appears that there are two schools of thought developing, as both critique, practice and advocacy.
One school is AI as an enhancement tool. That is very much how you have framed it and how I am also using AI although not in some of the more creative as way as you are. My own use has been in transcribing 19th and 18th century handwriting. And in copyediting suggestions for paragraphs I’ve already written. I also use it to find things in text different to what I have found.
The second school of thought is the replacement of humans and loss of skills. This is the worry. And it certainly exists, to greater or lesser degrees. I find it hard to see the copyediting profession being able to be sustained in the numbers it once was.
But maybe it will enhance the more active parts of our profession, we may spend more time in archives and travelling. Which also leads to major social changes.
Yes, it is interesting to think about the rebalancing of what is "special" about various knowledge work type jobs. My ability to proofread and line edit quickly -- once my prized job skill! -- is now functionally worthless. But maybe my ability to find unusual things in physical archives is more valuable because less machine-replicable? This is where I think the concept of "negative space around AI" is helpful because AI systems tend to be very bad at knowing what they don't know. They exist online and are fed a soup of web-scraped data, so they tend to conflate all knowledge with digitally available knowledge. So our job now is, in part, to think through the non-digitized, non-obvious knowledge or insights that we can uniquely access. That approach has always been productive of humanistic knowledge, the difference is that now that type knowledge is actually significantly more interesting and rare, not less.
Thank you for helping me put a finger on a weird feeling in my gut slash back of my mind I've had since early 2025 lol
Y'see, in many ways last year was one of the best of my of my professional life. Claude Code was the thing that broke through my AI skepticism, and that allowing me to leverage stacks outside my wheelhouse, combined with the time I saved with learning transcription flows with Gemini, automatic meeting notes processes etc etc etc
But I realise now that I miss the mindfulness that comes with writing up those notes, thinking of how to tweak my approach for a specific audience, how to keep things concise and deliver a call to action without sounding annoying or push even when sometimes it felt like busywork...
I think 2026 is gonna be my year of me trying to figure out how to balance that. But writing this out "cold turkey" without an AI chatbot tab in my peripheral vision and no Claude pinging me for approvals seems to have done wonders for my brain tbh...
Anyways, as you can tell conciseness was never my strength, but based on the fact you helped me with a personal breakthrough, you've earned a year's sub from me! 🫡
Thank you for subscribing and really genuinely pleased it was of help. I made much the same resolution for 2026. Back to writing research notes in an actual notebook for me...
Really pleased to see you talk about non-digitised sources. I have been saying this on repeat - AI can only access what’s accessible online. I keep thinking of the 100 miles of shelving just at the National Archive in the UK. It will probably get digitised some day but not anytime soon! The internet is not the world.
Yes very much agreed, I read this after my note on Gareth's comment but yes, I think about Margaret Mead's archives at the Library of Congress which are 530,000 items in total, of which only 31k are digitized. That proportion seems pretty typical for large archives like that. Also the oral histories sitting in people's brains that are waiting to be written down, but only if you pick up a phone or meet with them in person.
It seems to me that there is an ecosystem aspect to reading publics though. I.e. the profits that accrued to a publisher from the runaway success of stuff like "The Kingdom of Slender Swords" (good title, by the way!) might also make it possible for them to also publish, say, late period Henry James novels which maybe didn't sell a lot. These are not separate systems but are linked together through financially and socially in other words (same balance sheets, same editors, etc).
From talking to agents and editors, I gather that this has not changed much between 1910 and 2026. This implies that the bottom dropping out from the pulp novel market could still have lots of bad second order effects for literary publishing, bookstores, journalism, etc.
Interesting point, had not considered. It’s the same as people like dog groomers and chefs whose jobs are not mostly transferable to AI: their jobs may be at risk when their customers can't employ them any more due to being laid off.
‘The work is itself the point.’ Superb. What a tremendous exploration and explanation of the difference between AI writing and ‘my’ writing. I’m just reading The Singularity is Nearer by Ray Kurzweil and in a short section I’m going to post on my Notes later he makes it very clear how absolutely central to AI is language, analysed in a 500 dimensional ‘space’. But whilst it may fake feeling does AI feel? Or does it fake it till it makes it? I think your emphasis on personal style (that evolves of course) is spot on. ‘The frontlines of cognitive debt, in other words, are in the realms of both writing and coding.’ Thanks.
Not exactly on your topic, but maybe related: I took a quick try of your history simulator (a fascinating idea) and it broke down immediately. Since I'm in a book group on Zionism at my synagogue, I told the simulator to imagine I was a Zionist speaker in Russia in 1902. The simulator placed me in Novgorod, a place where Jews weren't allowed to live at the time (outside the "Pale of Settlement") and claimed I was devoutly religious, which is highly unlikely (Jewish religious authorities in Russia were opposed to Zionism at that time). If I'm getting nonsense that quickly, I wouldn't put much faith in it as a teaching tool, or as a generator of, say, historical fiction.
Of course, if you are training it on the Internet for a contested subject like Zionism, you might expect trouble. Maybe I should "play" it a bit and see if it descends into antisemitism. It wouldn't be the first time for AI.
Exactly right, the goal when using it in teaching isn’t to pretend that any of this is 100% accurate. It’s a simulacrum, after all. Instead it’s a novel way to encourage students to fact check and research their own answers to questions they pose about how and why it breaks. Thus getting students to perform historical research and reasoning (about historical anachronism, causality, and so forth) which is, ideally, more targeted and specific than if I had just had them write an essay about such and such time period or place. So in other words, I want students to employ precisely the type of thought process you applied here. The learning comes from figuring out what it gets wrong.
Fascinating! I'm an economist, so I guess I should try some economics on it! (Maybe my next try will be a member of the Federal Reserve Board in 1929). I do love the idea of learning by simulation, so (to be clear) I really respect what you have done. It's just...when I'm presented with something that off the rails, I do have to wonder. (And I see, "simulacrum" is different than "simulation," right?)
The risk to higher ed is not robots in the classroom, it's new companies that offer 80% of the value for 20% of the cost. Currently a four-year college degree is a reliable signal to employers, but I could see that changing rapidly.
I appreciate this post a lot and find it interesting because I found it almost immediately after I made a post about my new writing process — and realized we'd been writing through adjacent problems. Nice work Substack algorithm!
You're describing writing as an irreplaceable form of thinking — solitary perception tested against a reading public. I'd argue that's exactly what I'm trying to protect/resurface in my own newly created AI-assisted writing practice, not dissolve.
A few weeks ago, there was no space in my life for writing, today I can juuust sneak it in. And it feels amazing.
The comments on your post, this response, the exchange we might have — that's not replicable by a history simulator. The solitude-tested-against-a-readership you describe is precisely what makes this worth doing. Worth trying to sneak in.
Not sure what we are going to do with all the slop and sensationalized click-bait though. 🫠
The breadth of examples you've built vibe coding is extremely impressive and really helps show how AI can be a transformative tool that may help students or faculty really think about how this technology might actually be used in the classroom.
La deuda cognitiva que describes se asienta de forma diferente según quién la adquiere. Para muchas mujeres que entran tarde a la tecnología, delegar en IA representa la primera vez que no tienen que justificar cada paso de su proceso.
El problema es cuando eso se convierte en dependencia antes de que hayas tenido tiempo de confiar en tu propio razonamiento.
Your writing flows smoothly. Thank you. Notably absent in it is the LLM obsession with "It's not X, it's Y" "He didn't set out to A, he set out to B" I don't know if LLMs can even construct a paragraph without that format. Humans certainly have written that way, I've pick up books from my shelf and I see the style, but usually at the end of a paragraph. Mostly it's in popular nonfiction, an old trick of the pulp-master crank-a-books I guess. It's so bad now I have to skim essays, I'm just so tired of that LLM-style.
As an engineer in tech, it feels like I'm forced to live and breathe these tools and there's no going back.
But it feels so unfeasible for writers and artists to do the same with their craft. I can see it happening with software engineering and office work in general but it would be incredibly disheartening to see it happen to writing and art. I learn to draw on the side just for self-gratification and even that feels like a waste these days.
AI while exciting makes me depressed to think about my career and the things that I love.
"On the one hand, I feel a real sense of loss for the world of early 2000s New Yorker style long-form writing and criticism, the world of The New York Review of Books and Lapham’s Quarterly and the amazingly good Dennis Johnson novel I’m currently reading (Tree of Smoke)."
Is this because you think AI is capable (now or in the future) of writing essays that are publishable in these venues, making writers redundant? Or do you see them as likely casualties of the decline of intellectual culture abetted by AI?
I teach composition at a public university, and while the AI models can generate texts that are more competent than much of what my students are turning in, I have yet to read a text that I know is AI-generated that I can imagine reading in the New Yorker or NYRB. Let alone a great novel. The last time I played around with AI writing capabilities was about six months ago, and I was mostly unimpressed. (It did surprise me a few times.) But I understand they've made some big advances since then, so maybe I'm ignorant.
By the way, I really appreciate your writing on AI. It's difficult to get a handle on its impact on the humanities, because most humanists seem to dislike it reflexively, and people lauding its writing capabilities don't always come across as the most discerning readers. I've been reading your online writing since The Appendix, and I'm glad for your insights on this . . . sticky issue.
More the latter (not so much "likely casualties" as "crowded out from audience share in relative terms" - I am struck by how mass market the audience for NYRB/New Yorker type writers was up until recently, relative to today where that "middlebrow" world is seemingly being swamped by people's solo ChatGPT conversations, or by AI-written video "deep dives" etc. So it's not like those venues will go away, but they are for sure no longer part of a bigger cultural conversation in the same way, which bums me out because I love The New Yorker and I tend to love many of writers championed in venues like that in the 90s and early 2000s.
There are no AI safe jobs. If AI gets millions of people into unemployment, the money for plumbers, electricians, and even worse worse of surf instructors wont be there either. Even if there was, lots of people will try to get into such trades, bringing wages down there too.
As for historians and other academics, their job also depends on people having money to spend on college tuition and loans.
Time will tell, but there is a circularity here, since if you reject the premise that AI will fully replace most white collar work, then there is no grounds for assuming it will unemploy millions. It seems self-evident that some jobs will be lost due to AI (just as many professions were made obsolete during the advent of personal computing in the 1950s-1990s or the height of the machine age in the 1880s and 1890s). But what isn't clear yet is how many new jobs will created alongside that.
Enlightening read. Using AI as a tool, why not? But since “writing is thinking,” indeed, and allowing AI to do one's thinking instead is a disservice to both the writer and the readers.
My Dad died in Sept 2024 and I was the one to write his obituary. For many days, I could not produce a single word. I just couldn't get started. I really had no idea what to write. My kids and niece all suggested that I ought to use ChatGPT. Feed it the details of his life and let it do the work.
Perhaps that would have yielded something better, but I simply waited until the words, finally, came forth from my own fingers. Anything else would have felt as if I hadn't honored the memory and life of my father. I think that was an important choice for me personally.
A smart chap once said “For the things we have to learn before we can do them, we learn by doing them.” How many of my cognitive endeavors am I willing to outsource?
Thanks for the great article.
And one addendum: Cory Doctorow's enshittification of Internet platforms will leave no stone unturned:
https://www.fintechbrainfood.com/p/the-enshittification-of-chatgpt
Thanks for reading and commenting Mark, and sorry for your loss. I had a somewhat analogous experience in 2023 when my older daughter was in early infancy and obsessed with fairytales (she still is!). I asked ChatGPT to make up a custom fairytale for her. After reading the first line, I felt icky about it, turned off my phone, and made up a story of my own.
It appears that there are two schools of thought developing, as both critique, practice and advocacy.
One school is AI as an enhancement tool. That is very much how you have framed it and how I am also using AI although not in some of the more creative as way as you are. My own use has been in transcribing 19th and 18th century handwriting. And in copyediting suggestions for paragraphs I’ve already written. I also use it to find things in text different to what I have found.
The second school of thought is the replacement of humans and loss of skills. This is the worry. And it certainly exists, to greater or lesser degrees. I find it hard to see the copyediting profession being able to be sustained in the numbers it once was.
But maybe it will enhance the more active parts of our profession, we may spend more time in archives and travelling. Which also leads to major social changes.
Yes, it is interesting to think about the rebalancing of what is "special" about various knowledge work type jobs. My ability to proofread and line edit quickly -- once my prized job skill! -- is now functionally worthless. But maybe my ability to find unusual things in physical archives is more valuable because less machine-replicable? This is where I think the concept of "negative space around AI" is helpful because AI systems tend to be very bad at knowing what they don't know. They exist online and are fed a soup of web-scraped data, so they tend to conflate all knowledge with digitally available knowledge. So our job now is, in part, to think through the non-digitized, non-obvious knowledge or insights that we can uniquely access. That approach has always been productive of humanistic knowledge, the difference is that now that type knowledge is actually significantly more interesting and rare, not less.
Thank you for helping me put a finger on a weird feeling in my gut slash back of my mind I've had since early 2025 lol
Y'see, in many ways last year was one of the best of my of my professional life. Claude Code was the thing that broke through my AI skepticism, and that allowing me to leverage stacks outside my wheelhouse, combined with the time I saved with learning transcription flows with Gemini, automatic meeting notes processes etc etc etc
But I realise now that I miss the mindfulness that comes with writing up those notes, thinking of how to tweak my approach for a specific audience, how to keep things concise and deliver a call to action without sounding annoying or push even when sometimes it felt like busywork...
I think 2026 is gonna be my year of me trying to figure out how to balance that. But writing this out "cold turkey" without an AI chatbot tab in my peripheral vision and no Claude pinging me for approvals seems to have done wonders for my brain tbh...
Anyways, as you can tell conciseness was never my strength, but based on the fact you helped me with a personal breakthrough, you've earned a year's sub from me! 🫡
Thank you for subscribing and really genuinely pleased it was of help. I made much the same resolution for 2026. Back to writing research notes in an actual notebook for me...
Really pleased to see you talk about non-digitised sources. I have been saying this on repeat - AI can only access what’s accessible online. I keep thinking of the 100 miles of shelving just at the National Archive in the UK. It will probably get digitised some day but not anytime soon! The internet is not the world.
Yes very much agreed, I read this after my note on Gareth's comment but yes, I think about Margaret Mead's archives at the Library of Congress which are 530,000 items in total, of which only 31k are digitized. That proportion seems pretty typical for large archives like that. Also the oral histories sitting in people's brains that are waiting to be written down, but only if you pick up a phone or meet with them in person.
i'm not particularly worried for literary writing. trash is king; always has been. here is a list of the top 10 selling novels in the USA in 1910:
1. The Rosary by Florence L. Barclay
2. A Modern Chronicle by Winston Churchill
3. The Wild Olive by Anonymous (Basil King)
4. Max by Katherine Cecil Thurston
5. The Kingdom of Slender Swords by Hallie Erminie Rives
6. Simon the Jester by William J. Locke
7. Lord Loveland Discovers America by C. N. Williamson and A. M. Williamson
8. The Window at the White Cat by Mary Roberts Rinehart
9. Molly Make-Believe by Eleanor Hallowell Abbott
10. When a Man Marries by Mary Roberts Rinehart
the only one of these people I've ever heard of is Winston Churchill and only because he amusingly shares the name with Sir Winston Churchill
It seems to me that there is an ecosystem aspect to reading publics though. I.e. the profits that accrued to a publisher from the runaway success of stuff like "The Kingdom of Slender Swords" (good title, by the way!) might also make it possible for them to also publish, say, late period Henry James novels which maybe didn't sell a lot. These are not separate systems but are linked together through financially and socially in other words (same balance sheets, same editors, etc).
From talking to agents and editors, I gather that this has not changed much between 1910 and 2026. This implies that the bottom dropping out from the pulp novel market could still have lots of bad second order effects for literary publishing, bookstores, journalism, etc.
Interesting point, had not considered. It’s the same as people like dog groomers and chefs whose jobs are not mostly transferable to AI: their jobs may be at risk when their customers can't employ them any more due to being laid off.
‘The work is itself the point.’ Superb. What a tremendous exploration and explanation of the difference between AI writing and ‘my’ writing. I’m just reading The Singularity is Nearer by Ray Kurzweil and in a short section I’m going to post on my Notes later he makes it very clear how absolutely central to AI is language, analysed in a 500 dimensional ‘space’. But whilst it may fake feeling does AI feel? Or does it fake it till it makes it? I think your emphasis on personal style (that evolves of course) is spot on. ‘The frontlines of cognitive debt, in other words, are in the realms of both writing and coding.’ Thanks.
Not exactly on your topic, but maybe related: I took a quick try of your history simulator (a fascinating idea) and it broke down immediately. Since I'm in a book group on Zionism at my synagogue, I told the simulator to imagine I was a Zionist speaker in Russia in 1902. The simulator placed me in Novgorod, a place where Jews weren't allowed to live at the time (outside the "Pale of Settlement") and claimed I was devoutly religious, which is highly unlikely (Jewish religious authorities in Russia were opposed to Zionism at that time). If I'm getting nonsense that quickly, I wouldn't put much faith in it as a teaching tool, or as a generator of, say, historical fiction.
Of course, if you are training it on the Internet for a contested subject like Zionism, you might expect trouble. Maybe I should "play" it a bit and see if it descends into antisemitism. It wouldn't be the first time for AI.
Exactly right, the goal when using it in teaching isn’t to pretend that any of this is 100% accurate. It’s a simulacrum, after all. Instead it’s a novel way to encourage students to fact check and research their own answers to questions they pose about how and why it breaks. Thus getting students to perform historical research and reasoning (about historical anachronism, causality, and so forth) which is, ideally, more targeted and specific than if I had just had them write an essay about such and such time period or place. So in other words, I want students to employ precisely the type of thought process you applied here. The learning comes from figuring out what it gets wrong.
Fascinating! I'm an economist, so I guess I should try some economics on it! (Maybe my next try will be a member of the Federal Reserve Board in 1929). I do love the idea of learning by simulation, so (to be clear) I really respect what you have done. It's just...when I'm presented with something that off the rails, I do have to wonder. (And I see, "simulacrum" is different than "simulation," right?)
The risk to higher ed is not robots in the classroom, it's new companies that offer 80% of the value for 20% of the cost. Currently a four-year college degree is a reliable signal to employers, but I could see that changing rapidly.
I appreciate this post a lot and find it interesting because I found it almost immediately after I made a post about my new writing process — and realized we'd been writing through adjacent problems. Nice work Substack algorithm!
You're describing writing as an irreplaceable form of thinking — solitary perception tested against a reading public. I'd argue that's exactly what I'm trying to protect/resurface in my own newly created AI-assisted writing practice, not dissolve.
A few weeks ago, there was no space in my life for writing, today I can juuust sneak it in. And it feels amazing.
The comments on your post, this response, the exchange we might have — that's not replicable by a history simulator. The solitude-tested-against-a-readership you describe is precisely what makes this worth doing. Worth trying to sneak in.
Not sure what we are going to do with all the slop and sensationalized click-bait though. 🫠
— Amber
The breadth of examples you've built vibe coding is extremely impressive and really helps show how AI can be a transformative tool that may help students or faculty really think about how this technology might actually be used in the classroom.
La deuda cognitiva que describes se asienta de forma diferente según quién la adquiere. Para muchas mujeres que entran tarde a la tecnología, delegar en IA representa la primera vez que no tienen que justificar cada paso de su proceso.
El problema es cuando eso se convierte en dependencia antes de que hayas tenido tiempo de confiar en tu propio razonamiento.
Your writing flows smoothly. Thank you. Notably absent in it is the LLM obsession with "It's not X, it's Y" "He didn't set out to A, he set out to B" I don't know if LLMs can even construct a paragraph without that format. Humans certainly have written that way, I've pick up books from my shelf and I see the style, but usually at the end of a paragraph. Mostly it's in popular nonfiction, an old trick of the pulp-master crank-a-books I guess. It's so bad now I have to skim essays, I'm just so tired of that LLM-style.
As an engineer in tech, it feels like I'm forced to live and breathe these tools and there's no going back.
But it feels so unfeasible for writers and artists to do the same with their craft. I can see it happening with software engineering and office work in general but it would be incredibly disheartening to see it happen to writing and art. I learn to draw on the side just for self-gratification and even that feels like a waste these days.
AI while exciting makes me depressed to think about my career and the things that I love.
"On the one hand, I feel a real sense of loss for the world of early 2000s New Yorker style long-form writing and criticism, the world of The New York Review of Books and Lapham’s Quarterly and the amazingly good Dennis Johnson novel I’m currently reading (Tree of Smoke)."
Is this because you think AI is capable (now or in the future) of writing essays that are publishable in these venues, making writers redundant? Or do you see them as likely casualties of the decline of intellectual culture abetted by AI?
I teach composition at a public university, and while the AI models can generate texts that are more competent than much of what my students are turning in, I have yet to read a text that I know is AI-generated that I can imagine reading in the New Yorker or NYRB. Let alone a great novel. The last time I played around with AI writing capabilities was about six months ago, and I was mostly unimpressed. (It did surprise me a few times.) But I understand they've made some big advances since then, so maybe I'm ignorant.
By the way, I really appreciate your writing on AI. It's difficult to get a handle on its impact on the humanities, because most humanists seem to dislike it reflexively, and people lauding its writing capabilities don't always come across as the most discerning readers. I've been reading your online writing since The Appendix, and I'm glad for your insights on this . . . sticky issue.
More the latter (not so much "likely casualties" as "crowded out from audience share in relative terms" - I am struck by how mass market the audience for NYRB/New Yorker type writers was up until recently, relative to today where that "middlebrow" world is seemingly being swamped by people's solo ChatGPT conversations, or by AI-written video "deep dives" etc. So it's not like those venues will go away, but they are for sure no longer part of a bigger cultural conversation in the same way, which bums me out because I love The New Yorker and I tend to love many of writers championed in venues like that in the 90s and early 2000s.
Also I meant to say thank you for reading since the Appendix, that is so nice to hear!
There are no AI safe jobs. If AI gets millions of people into unemployment, the money for plumbers, electricians, and even worse worse of surf instructors wont be there either. Even if there was, lots of people will try to get into such trades, bringing wages down there too.
As for historians and other academics, their job also depends on people having money to spend on college tuition and loans.
Time will tell, but there is a circularity here, since if you reject the premise that AI will fully replace most white collar work, then there is no grounds for assuming it will unemploy millions. It seems self-evident that some jobs will be lost due to AI (just as many professions were made obsolete during the advent of personal computing in the 1950s-1990s or the height of the machine age in the 1880s and 1890s). But what isn't clear yet is how many new jobs will created alongside that.
Enlightening read. Using AI as a tool, why not? But since “writing is thinking,” indeed, and allowing AI to do one's thinking instead is a disservice to both the writer and the readers.