It's actually more important as a skill to learning than the author realizes. It connects the brain to the CNS for deeper processing and memory retention of information.
Interesting comment. I’ve always found that writing something down helps me remember (even when I never look at notes again) but does the writing have to be in cursive?
There will be generations of children taught by generations of adults who did not have cursive as a core tool for learning. We’re either screwed or well on our way to the future imagined by a Star Trek Discovery episode.
Writing doesn't have to be in cursive, but cursive can allow for faster, more fluid writing. Research shows taking notes by hand—especially in cursive—enhances comprehension and helps the brain retain information better than typing on a keyboard. Also isn't there value in learning something because of its beauty?
Christina, are you saying things can have intrinsic value all on their own and that we shouldn't just completely dump something as soon as something with more utility comes along? Sight unseen! (and I also agree, there's value with writing by hand, both in terms of its cognition-boosting ability, and its virtue as a beauty in-and-of itself). True story: yesterday I received a hand-written letter in the mail. The legibility was quite low, but as it turns out, this person was in a serious auto collision, and is now sending hand-written letters to people to help re-learn fine motor skills! :-)
Ontario also recently (2 years ago lol) re-introduced cursive into the curriculum alongside a renewed focus on phonics! Sometimes you don't know what you're missing until a generation grows up without it and you see the difference yourself.
I agree. The "friction" of writing cursive is a great mental and physical exercise to slow the brain down and make one chew on their thoughts (and writings).
I have just read this article and wanted to point out exactly the same thing. As one study was already mentioned in the comments show, there is a substantial body of research on this topic. What I find interesting is how accurate this article is in many respects (thanks for sharing these thoughts), while at the same time failing to address one of the most important aspect when it comes to learning. Writing by hand is a whole different game for your brain – more effort, more outcome.
I see where you are coming from, but you can distinguish who is thinking on their own with low to medium support from AI and who is relying on AI to produce enormous amount of content that regular person wouldn’t be able to come up with. It shows it in the depth. It shows in the voice. It shows in the quality. AI is great to start and I see many starting but then giving up because it’s not sustainable or they didn’t get the expected results. Let’s also keep in mind that in the internet you don’t see the whole story of other people and let’s not assume we know it all - for instance for me the turning point to start writing was sabbatical leave, not the fact that I had suddenly access to the tool and could produce unlimited content. I wouldn’t speak if I had nothing to say.
The tricky thing is not just the LinkedIn algorithms but those we subconsciously live by: what we read and what we filter out. Usually the AI posts are properly attention grabbing, “feeling like they’re saying something” and optimized to keep a readers attention.
I don’t think the core problem is AI-generated writing.
The deeper shift is that AI dramatically lowers the cost of producing “plausible” output.
And once plausibility becomes cheap, people and organizations stop investing in the slower processes that used to build judgment: rewriting, defending ideas, reading deeply, receiving criticism, deciding when something is actually “done.”
Those activities were never just about producing text.
They were also developmental processes for building discernment, cognitive resilience, and responsibility.
The risk is not simply loss of authenticity.
It’s the erosion of the structures that produce judgment in the first place.
We may end up generating increasingly polished outputs with increasingly underdeveloped evaluative capacity behind them.
I've fallen into bad habits a number of times due to time-crunch and sheer over commitment of letting AI do the hard part. Or the rote part, at least. I realized fairly recently though that I could bake in my own "learning" mode pretty easily on all the major product platforms, though it does come at a bit of a token cost.
I've amended the instructions/memory of most of my platforms to "communicate in a Socratic method to refine confidence and assist my own output rather than supply finalized content". It usually results in answers that get right to the point if it's a simple request or already something inclined to ideation. As soon as it looks like I'm reaching for a quick draft of a post, article, or training aid at work though, it starts peppering me back with questions and assistance like outlines rather than a finalized piece right away.
These instructions are excellent. I've done the same thing and try building that into specific "dialogic workflows." In case you are interested, I just posted today about how I am working to constrain Claude Cowork (my preferred platform now, tbh) to support academic and intellectual work. The foundation of that is a system that establishes a collaboration model, a "relation" model that governs how I interact with the tool and the tool interacts with me, and a cognitive budget allocation to protect the necessary deliberation and review that kind of work requires. It's interesting to see the varied responses here as people have started trying to wrangle these tools, which we will not be able to evade forever, into supporting the kind of thinking they want to do. https://mivonnewf.substack.com/p/seatbelts-for-pod-racers-cognitive
I just so instructed my workspace in ChatGPT. Good suggestion! Alas, I have low confidence in the veracity of ChatGPT's reply to my instruction, which read simply: "Remembered."
I've definitely had to call out a few different chats that let the instruction slip. It's always hilariously contrite for some reason, regardless of the model. Something to do with the influence of "Socratic method" I'm guessing!
Not at all! Happy to compare notes any time. I firmly believe we haven't uncovered a fraction of the best practices and amazing opportunities available with just the AI we have today, and knowledge sharing is the only way to even get started on diving deeper!
Do you follow the work of Rebecca Winthrop, who is leading the Brookings Global Task Force on AI and Education? They are examining what Gen AI means for students’ learning and development. She has coined the term "cognitive stunting" and writes that it is often the result of wide AI use. While narrow AI use that is targeted and intentionally deployed by teachers can support thinking and learning. Here's the link to the article where she discusses this - https://www.brookings.edu/articles/is-it-time-to-measure-cognitive-stunting/
The Turkey vs. Taipei contrast is doing a lot of work here, and I think it generalizes beyond education.
Same technology. Two deployments. Categorically different outcomes. The variable isn't AI capability or even user intentionality as a disposition. It's whether the deployment scaffolds the effortful work or short-circuits it. The Anthropic programmer study you cite points the same direction: the ones who asked the AI to explain, or who delegated only parts, kept the loop intact. The ones who surrendered the whole task lost the learning that would have come from doing it.
This is why I think "be intentional" can only carry so much weight as advice. Intentionality is a disposition. What actually produces different outcomes is practice, a learned habit of using AI as friction against your first thought rather than as an amplifier of it. Most people, left to defaults, do the opposite, because the defaults are built for frictionless output.
Your final point is the one I keep coming back to in my own writing: the defaults are being set right now, by people optimizing for use rather than for what use does to us over time. Reversing them later, after a generation has built habits around them, will be much harder than getting them right now.
Absolutely! And drifting away from "being intentional" is likely inevitable under high cognitive load, or when we are stressed, etc. I actually think that teaching students how to constrain these tools to support the kind of thinking they want to do is a critical imperative right now, as these tools are being shipped for engagement and the lowest possible friction. And I think its important to make those strategic decisions about how we want the tool to support our way of thinking before we start using them, when we are clear headed and can design guardrails that support us when, not if, we drift away from those first principles. We need to make early decisions about the collaboration model we want to use, how to relate to the tool and how the tool should "relate" to us, and how to allocate our cognitive budget to support deliberation. And we need to bake them into our tools as foundational governing principles. I just posted about this, describing my version of this as Cognitive Fitness Protocols: https://mivonnewf.substack.com/p/seatbelts-for-pod-racers-cognitive
I joined a Substack that promotes writers on the subject of AI. Within a couple of weeks of joining, I was disillusioned by the "article mill". The truly disheartening thing is how passionate people are about getting their thinking about there, but as you cite, it's obvious when you see "the script": Claude as a 'ghost writer' is the equivalent of the Haunted Mansion at Disneyworld. It really is off-putting once you recognise it. I've gotten to the point where if I hear a YouTube video narration begins and I identify the style I cut and run, because I now feel like the content can't be trusted (whether that's fair or not).
I think it does comes down to us, but not entirely, and maybe not even mostly. I think we have more agency than we tend to use, and perhaps the right move is to use AI to push our thinking rather than shortcut it. But that's a heavy burden to put on a student or a professional when the shortcut gets it done, turned in, and meets the deadline. And the "what to keep human" question is harder than even that, because the choice often isn't made at the moment of use, the choice is made upstream, in the architecture. Because a frictionless agent has already decided what to hand you before you get a vote on it. Intentionality is a user-level fix for something that looks increasingly settled at the system level, by people optimizing for frictionless easy buttons.
So "choosing to stay human" may be as much an architecture question as a discipline one. A question about what gets built, and not only about how we use it.
A big problem with AI-written content beyond just the problem of repetitive writing patterns that become annoying, is that I have no idea of the person even reviewed what they posted. Since so much content is just created for engagement or SEO hacking on full autopilot, I tend to expect low quality in anything that sounds AI.
We must still keep the things that define us as human even if we are using AI tools. Most of us don't want to talk to bots or people that sound like bots. We want to have conversations with other humans. You can use AI, but don't make it the middleman between conversations we are having.
Authenticity needs to be a cultural value in the midst of this AI world in order to keep our sanity and values for life. More thoughts on that I wrote a while back here. FYI.
I feel this is even worse in AI driven software development. As a vibe coding Product Manager who used to be a coder, I've personally been struggling with what I call comprehension debt (similar to cognitive surrender) when developing complex applications needing huge amounts of rapid planning and coding.
I literally feel selfish when I spend a LOT more time understanding the LLM responses to my complex questions and architectural and coding problems. The speed at which these tools spout intelligent solutions for everything, at inhuman speeds, make it impossible to actually claim you can comprehend it all even if it does work. And the makers of the tech actively want you to simply ignore all the intermediate thinking and analysis.
In real life, if I were in a class or a meeting with colleagues I would not say "sorry I don't understand" 10 times in 30 mins for fear of appearing stupid. In a way, this may be one of the reason why solutions to this could be hard. Our social interactions do not reward incessant questioning and yet these inhuman responses require questioning for true understanding as opposed to cognitive surrender. Most of our society rewards appearing smart and quick on the draw, but that's compared to other humans and there's the rub. We are not dealing with other human intelligence!
But on introspection, there's another reason to have reached this point after 7-8 months of this coding mode. The pressure to show productivity is another reason I don't give myself permission to slow down. If the highest value our society and corporations want is productivity at all costs, then even acknowledging "comprehension debt" seems selfish. If I said to a manager I need to slow down to reduce my comprehension debt, I'd be out of a job in a month.
I'm not a software developer, but I feel the same squeeze every day. "Comprehension debt" is a good term, and I think it names something the other two miss. "Cognitive debt" and "cognitive surrender" put the problem in the user. I tihnk yours puts it where the speed actually comes from: the tool spouts an architecture-grade answer in seconds, and, as you say, its makers would prefer you skip the intermediate reasoning entirely.
That's not you surrendering.
That's the system handing you a finished thing and hoping you don't look closely, and which we generally don't. (The MIT Media Lab "Your Brain on ChatGPT" study, Kosmyna et al. 2025, formalized "cognitive debt" for this — but the phrase has older roots in neuroscience too.)
Which is why I keep landing where you do at the end: the productivity treadmill means we won't win this foot race, so the fix is perhaps more about architecture and less about user willpower. The one piece I'd add, perhaps alongside better-built tools, we may need to deliberately train knowledge formation in students and knowledge maintenance in professionals. Because once you pull the AI away, what's left?
> "That style took a lot of super annoying work to get to: good teachers and rewrites and mean online comments all contributed. If the AI does fine writing, I could skip all of that, but I would have done so the cost of giving up something that has turned out to be very important to my career and my happiness."
Ironic that there's a grammatical error here, that an AI wouldn't have made :)
Should be: "I would have done so *at* the cost of"
Another unfortunate irony is that the comments on this piece aren't particularly substantive, despite the fact that they're presumably (mostly) produced by humans.
I have noticed a flood of AI-written posts and scripts recently, especially on YouTube, and it is not just that the style becomes tedious once you recognise it.
Because AI writing is token-led, it tends towards the most predictable next word, the most familiar phrase, and the most common shape of an argument. My concern is that, used lazily, it could narrow the range of human thought and debate.
That matters, because we need the widest possible range of views, ideas and opinions to inform the wisdom of the crowd. That variety has helped human beings adapt since prehistory.
You may find this relevant, Ethan. I wrote about a related idea here: Why We’re All Wrong About Everything All the Time.
I've been exploring the institutional version of the question you explore here — what happens when the surrender is organisational? A manager who stops walking the factory floor because the dashboard in his office is green. A doctor whose diagnostic instinct was built across thousands of patients, now replaced by a resident who validates AI suggestions. The drift is gradual. You feel competent. And then one day the system fails and the muscle that used to catch the error has atrophied.
Your point about defaults being set without planning is a theme I wish we knew how to action more effectively. At the individual level, intentionality is possible — your learning modes, your "push me to think" suggestion. At the institutional level, the defaults are hardening fast and nobody is steering. That may be a harder problem.
At the AI Squared Summit this year in Orlando, a conference focused on AI in higher ed, two philosophers presented a poster looking at this institutional level question. I don't think they have published the work yet, but it may be of interest to you when they do: Rodrigo Borges and Panagiotis Saranteas. “Beyond Blind Trust: AI Sponsors and Virtuous Communities As Safeguards for Epistemic Agency in the Age of Cognitive Delegation,” poster presented at the AI Squared Summit, Orlando, FL, 2026. They argue that at the institutional level, you need both an "AI Sponsor," an agent that monitors human use to ensure that skills don't atrophy and that humans are providing active consent for decisions that need to remain in human hands, AND a community of virtuous delegators who can keep themselves accountable for responsibly constraining their delegation. I found their "AI sponsor" proposal fascinating, as it was an institution level version of how I had started constraining Claude to support my own thinking -- in my case, I constrain it by establishing a collaboration model, a "relational" model that governs how I interact with the tool and how it interacts with me, and a cognitive budget allocation meant to ensure a human scale for deliberation and review. I just wrote this up, if you are interested: https://mivonnewf.substack.com/p/seatbelts-for-pod-racers-cognitive
Thank you for the Borges and Saranteas reference. They appear to be working in the terrain that I've been circling from the narrative side. The two-level architecture idea is excellent: the AI Sponsor as individual safeguard, the community of virtuous delegators as institutional safeguard. That maps onto a distinction I keep returning to — individual intentionality isn't enough if the institutional defaults are pushing the other way. Will track their work. My shorter take on what atrophies when the loop goes empty, if you're interested: https://rajeshachanta.substack.com/p/when-the-signal-drops
Will also read your piece - seatbelts for pod racers is intriguing.
Agreed with most of this but I'm glad my kids are still learning cursive in their school!
my AI proofreader warned me that my views on cursive would be the most controversial aspect of the post 😅.
Your AI proof reader was right! 😂
Meh, put me solidly in the no cursive camp, thank you very much.
-Ernie, on behalf of left-handers everywhere
It's actually more important as a skill to learning than the author realizes. It connects the brain to the CNS for deeper processing and memory retention of information.
Interesting comment. I’ve always found that writing something down helps me remember (even when I never look at notes again) but does the writing have to be in cursive?
Here is some research on why the pen is mightier than the (key)board.
https://www.frontiersin.org/news/2024/01/26/writing-by-hand-increase-brain-connectivity-typing
There will be generations of children taught by generations of adults who did not have cursive as a core tool for learning. We’re either screwed or well on our way to the future imagined by a Star Trek Discovery episode.
Writing doesn't have to be in cursive, but cursive can allow for faster, more fluid writing. Research shows taking notes by hand—especially in cursive—enhances comprehension and helps the brain retain information better than typing on a keyboard. Also isn't there value in learning something because of its beauty?
Christina, are you saying things can have intrinsic value all on their own and that we shouldn't just completely dump something as soon as something with more utility comes along? Sight unseen! (and I also agree, there's value with writing by hand, both in terms of its cognition-boosting ability, and its virtue as a beauty in-and-of itself). True story: yesterday I received a hand-written letter in the mail. The legibility was quite low, but as it turns out, this person was in a serious auto collision, and is now sending hand-written letters to people to help re-learn fine motor skills! :-)
That’s beautiful. And now I’m wishing a random person much joy in the writing and recovery.
That makes sense. Sounds like I find that I think more clearly when I write or draw in a notebook.
i share that sense too....... it is also beautiful.....cursive. it flows.... and the brain seems to flow differently too...fluid....
Ontario also recently (2 years ago lol) re-introduced cursive into the curriculum alongside a renewed focus on phonics! Sometimes you don't know what you're missing until a generation grows up without it and you see the difference yourself.
I agree. The "friction" of writing cursive is a great mental and physical exercise to slow the brain down and make one chew on their thoughts (and writings).
then you should check this out https://tinaaustin.substack.com/p/dont-call-it-ai-cognitive-surrender?r=4a98uc&utm_campaign=post-expanded-share&utm_medium=web
I have just read this article and wanted to point out exactly the same thing. As one study was already mentioned in the comments show, there is a substantial body of research on this topic. What I find interesting is how accurate this article is in many respects (thanks for sharing these thoughts), while at the same time failing to address one of the most important aspect when it comes to learning. Writing by hand is a whole different game for your brain – more effort, more outcome.
Is it not called running writting anymore?
You had me at "...meaning-shaped attention vampires that take mental effort to decode..."
‘something-shaped’ …. But it isn’t meaning.
One thing I really grapple with is that I can see so many technical professionals now pumping out substacks, essays and LinkedIn content.
They are clearly enthused by the prospect that their thoughts, feelings and emotions can now be translated into beautiful prose.
In a way, I am all here for it.
But as I start to consume the content, this feeling of dread creeps upon me:
“Oh no, this [actuary, accountant, scientist etc] has outsourced their mind to AI”
They have slid down a rabbit hole I fear might be impossible to climb back out of.
I see where you are coming from, but you can distinguish who is thinking on their own with low to medium support from AI and who is relying on AI to produce enormous amount of content that regular person wouldn’t be able to come up with. It shows it in the depth. It shows in the voice. It shows in the quality. AI is great to start and I see many starting but then giving up because it’s not sustainable or they didn’t get the expected results. Let’s also keep in mind that in the internet you don’t see the whole story of other people and let’s not assume we know it all - for instance for me the turning point to start writing was sabbatical leave, not the fact that I had suddenly access to the tool and could produce unlimited content. I wouldn’t speak if I had nothing to say.
The tricky thing is not just the LinkedIn algorithms but those we subconsciously live by: what we read and what we filter out. Usually the AI posts are properly attention grabbing, “feeling like they’re saying something” and optimized to keep a readers attention.
Meanwhile “real” posts go unnoticed.
I don’t think the core problem is AI-generated writing.
The deeper shift is that AI dramatically lowers the cost of producing “plausible” output.
And once plausibility becomes cheap, people and organizations stop investing in the slower processes that used to build judgment: rewriting, defending ideas, reading deeply, receiving criticism, deciding when something is actually “done.”
Those activities were never just about producing text.
They were also developmental processes for building discernment, cognitive resilience, and responsibility.
The risk is not simply loss of authenticity.
It’s the erosion of the structures that produce judgment in the first place.
We may end up generating increasingly polished outputs with increasingly underdeveloped evaluative capacity behind them.
I've fallen into bad habits a number of times due to time-crunch and sheer over commitment of letting AI do the hard part. Or the rote part, at least. I realized fairly recently though that I could bake in my own "learning" mode pretty easily on all the major product platforms, though it does come at a bit of a token cost.
I've amended the instructions/memory of most of my platforms to "communicate in a Socratic method to refine confidence and assist my own output rather than supply finalized content". It usually results in answers that get right to the point if it's a simple request or already something inclined to ideation. As soon as it looks like I'm reaching for a quick draft of a post, article, or training aid at work though, it starts peppering me back with questions and assistance like outlines rather than a finalized piece right away.
These instructions are excellent. I've done the same thing and try building that into specific "dialogic workflows." In case you are interested, I just posted today about how I am working to constrain Claude Cowork (my preferred platform now, tbh) to support academic and intellectual work. The foundation of that is a system that establishes a collaboration model, a "relation" model that governs how I interact with the tool and the tool interacts with me, and a cognitive budget allocation to protect the necessary deliberation and review that kind of work requires. It's interesting to see the varied responses here as people have started trying to wrangle these tools, which we will not be able to evade forever, into supporting the kind of thinking they want to do. https://mivonnewf.substack.com/p/seatbelts-for-pod-racers-cognitive
I just so instructed my workspace in ChatGPT. Good suggestion! Alas, I have low confidence in the veracity of ChatGPT's reply to my instruction, which read simply: "Remembered."
I've definitely had to call out a few different chats that let the instruction slip. It's always hilariously contrite for some reason, regardless of the model. Something to do with the influence of "Socratic method" I'm guessing!
I like the instruction you use. I hope you don’t mind that I plan to use it in Claude to assist me in the law school class I teach. Thanks!
Not at all! Happy to compare notes any time. I firmly believe we haven't uncovered a fraction of the best practices and amazing opportunities available with just the AI we have today, and knowledge sharing is the only way to even get started on diving deeper!
Do you follow the work of Rebecca Winthrop, who is leading the Brookings Global Task Force on AI and Education? They are examining what Gen AI means for students’ learning and development. She has coined the term "cognitive stunting" and writes that it is often the result of wide AI use. While narrow AI use that is targeted and intentionally deployed by teachers can support thinking and learning. Here's the link to the article where she discusses this - https://www.brookings.edu/articles/is-it-time-to-measure-cognitive-stunting/
Thanks!
The Turkey vs. Taipei contrast is doing a lot of work here, and I think it generalizes beyond education.
Same technology. Two deployments. Categorically different outcomes. The variable isn't AI capability or even user intentionality as a disposition. It's whether the deployment scaffolds the effortful work or short-circuits it. The Anthropic programmer study you cite points the same direction: the ones who asked the AI to explain, or who delegated only parts, kept the loop intact. The ones who surrendered the whole task lost the learning that would have come from doing it.
This is why I think "be intentional" can only carry so much weight as advice. Intentionality is a disposition. What actually produces different outcomes is practice, a learned habit of using AI as friction against your first thought rather than as an amplifier of it. Most people, left to defaults, do the opposite, because the defaults are built for frictionless output.
Your final point is the one I keep coming back to in my own writing: the defaults are being set right now, by people optimizing for use rather than for what use does to us over time. Reversing them later, after a generation has built habits around them, will be much harder than getting them right now.
Absolutely! And drifting away from "being intentional" is likely inevitable under high cognitive load, or when we are stressed, etc. I actually think that teaching students how to constrain these tools to support the kind of thinking they want to do is a critical imperative right now, as these tools are being shipped for engagement and the lowest possible friction. And I think its important to make those strategic decisions about how we want the tool to support our way of thinking before we start using them, when we are clear headed and can design guardrails that support us when, not if, we drift away from those first principles. We need to make early decisions about the collaboration model we want to use, how to relate to the tool and how the tool should "relate" to us, and how to allocate our cognitive budget to support deliberation. And we need to bake them into our tools as foundational governing principles. I just posted about this, describing my version of this as Cognitive Fitness Protocols: https://mivonnewf.substack.com/p/seatbelts-for-pod-racers-cognitive
I joined a Substack that promotes writers on the subject of AI. Within a couple of weeks of joining, I was disillusioned by the "article mill". The truly disheartening thing is how passionate people are about getting their thinking about there, but as you cite, it's obvious when you see "the script": Claude as a 'ghost writer' is the equivalent of the Haunted Mansion at Disneyworld. It really is off-putting once you recognise it. I've gotten to the point where if I hear a YouTube video narration begins and I identify the style I cut and run, because I now feel like the content can't be trusted (whether that's fair or not).
I think it does comes down to us, but not entirely, and maybe not even mostly. I think we have more agency than we tend to use, and perhaps the right move is to use AI to push our thinking rather than shortcut it. But that's a heavy burden to put on a student or a professional when the shortcut gets it done, turned in, and meets the deadline. And the "what to keep human" question is harder than even that, because the choice often isn't made at the moment of use, the choice is made upstream, in the architecture. Because a frictionless agent has already decided what to hand you before you get a vote on it. Intentionality is a user-level fix for something that looks increasingly settled at the system level, by people optimizing for frictionless easy buttons.
So "choosing to stay human" may be as much an architecture question as a discipline one. A question about what gets built, and not only about how we use it.
A big problem with AI-written content beyond just the problem of repetitive writing patterns that become annoying, is that I have no idea of the person even reviewed what they posted. Since so much content is just created for engagement or SEO hacking on full autopilot, I tend to expect low quality in anything that sounds AI.
We must still keep the things that define us as human even if we are using AI tools. Most of us don't want to talk to bots or people that sound like bots. We want to have conversations with other humans. You can use AI, but don't make it the middleman between conversations we are having.
Authenticity needs to be a cultural value in the midst of this AI world in order to keep our sanity and values for life. More thoughts on that I wrote a while back here. FYI.
https://www.mindprison.cc/p/make-authenticity-great-again
I feel this is even worse in AI driven software development. As a vibe coding Product Manager who used to be a coder, I've personally been struggling with what I call comprehension debt (similar to cognitive surrender) when developing complex applications needing huge amounts of rapid planning and coding.
I literally feel selfish when I spend a LOT more time understanding the LLM responses to my complex questions and architectural and coding problems. The speed at which these tools spout intelligent solutions for everything, at inhuman speeds, make it impossible to actually claim you can comprehend it all even if it does work. And the makers of the tech actively want you to simply ignore all the intermediate thinking and analysis.
In real life, if I were in a class or a meeting with colleagues I would not say "sorry I don't understand" 10 times in 30 mins for fear of appearing stupid. In a way, this may be one of the reason why solutions to this could be hard. Our social interactions do not reward incessant questioning and yet these inhuman responses require questioning for true understanding as opposed to cognitive surrender. Most of our society rewards appearing smart and quick on the draw, but that's compared to other humans and there's the rub. We are not dealing with other human intelligence!
But on introspection, there's another reason to have reached this point after 7-8 months of this coding mode. The pressure to show productivity is another reason I don't give myself permission to slow down. If the highest value our society and corporations want is productivity at all costs, then even acknowledging "comprehension debt" seems selfish. If I said to a manager I need to slow down to reduce my comprehension debt, I'd be out of a job in a month.
I'm not a software developer, but I feel the same squeeze every day. "Comprehension debt" is a good term, and I think it names something the other two miss. "Cognitive debt" and "cognitive surrender" put the problem in the user. I tihnk yours puts it where the speed actually comes from: the tool spouts an architecture-grade answer in seconds, and, as you say, its makers would prefer you skip the intermediate reasoning entirely.
That's not you surrendering.
That's the system handing you a finished thing and hoping you don't look closely, and which we generally don't. (The MIT Media Lab "Your Brain on ChatGPT" study, Kosmyna et al. 2025, formalized "cognitive debt" for this — but the phrase has older roots in neuroscience too.)
Which is why I keep landing where you do at the end: the productivity treadmill means we won't win this foot race, so the fix is perhaps more about architecture and less about user willpower. The one piece I'd add, perhaps alongside better-built tools, we may need to deliberately train knowledge formation in students and knowledge maintenance in professionals. Because once you pull the AI away, what's left?
> "That style took a lot of super annoying work to get to: good teachers and rewrites and mean online comments all contributed. If the AI does fine writing, I could skip all of that, but I would have done so the cost of giving up something that has turned out to be very important to my career and my happiness."
Ironic that there's a grammatical error here, that an AI wouldn't have made :)
Should be: "I would have done so *at* the cost of"
Another unfortunate irony is that the comments on this piece aren't particularly substantive, despite the fact that they're presumably (mostly) produced by humans.
Those are my mean online comments for the day
Interesting post Ethan, thank you.
I have noticed a flood of AI-written posts and scripts recently, especially on YouTube, and it is not just that the style becomes tedious once you recognise it.
Because AI writing is token-led, it tends towards the most predictable next word, the most familiar phrase, and the most common shape of an argument. My concern is that, used lazily, it could narrow the range of human thought and debate.
That matters, because we need the widest possible range of views, ideas and opinions to inform the wisdom of the crowd. That variety has helped human beings adapt since prehistory.
You may find this relevant, Ethan. I wrote about a related idea here: Why We’re All Wrong About Everything All the Time.
What keeps getting missed in these conversations is that “staying human” is not mainly about preserving inefficiency.
It is about preserving stakes.
A system can generate fluent output, but it cannot generate why this matters to this person, at this moment, under these consequences.
That layer is not style. It is the structure that gives writing, judgment, and responsibility their weight.
I've been exploring the institutional version of the question you explore here — what happens when the surrender is organisational? A manager who stops walking the factory floor because the dashboard in his office is green. A doctor whose diagnostic instinct was built across thousands of patients, now replaced by a resident who validates AI suggestions. The drift is gradual. You feel competent. And then one day the system fails and the muscle that used to catch the error has atrophied.
Your point about defaults being set without planning is a theme I wish we knew how to action more effectively. At the individual level, intentionality is possible — your learning modes, your "push me to think" suggestion. At the institutional level, the defaults are hardening fast and nobody is steering. That may be a harder problem.
Two adjacent pieces if you're interested — one on what atrophies when we delegate too much: https://rajeshachanta.substack.com/p/when-the-signal-drops. And one on who decides whether the AI transition might lead to a three-day work week or 40% unemployment: https://rajeshachanta.substack.com/p/the-doorman-and-deepseek (I paraphrase your taming instinct observation as I craft my argument)
At the AI Squared Summit this year in Orlando, a conference focused on AI in higher ed, two philosophers presented a poster looking at this institutional level question. I don't think they have published the work yet, but it may be of interest to you when they do: Rodrigo Borges and Panagiotis Saranteas. “Beyond Blind Trust: AI Sponsors and Virtuous Communities As Safeguards for Epistemic Agency in the Age of Cognitive Delegation,” poster presented at the AI Squared Summit, Orlando, FL, 2026. They argue that at the institutional level, you need both an "AI Sponsor," an agent that monitors human use to ensure that skills don't atrophy and that humans are providing active consent for decisions that need to remain in human hands, AND a community of virtuous delegators who can keep themselves accountable for responsibly constraining their delegation. I found their "AI sponsor" proposal fascinating, as it was an institution level version of how I had started constraining Claude to support my own thinking -- in my case, I constrain it by establishing a collaboration model, a "relational" model that governs how I interact with the tool and how it interacts with me, and a cognitive budget allocation meant to ensure a human scale for deliberation and review. I just wrote this up, if you are interested: https://mivonnewf.substack.com/p/seatbelts-for-pod-racers-cognitive
Thank you for the Borges and Saranteas reference. They appear to be working in the terrain that I've been circling from the narrative side. The two-level architecture idea is excellent: the AI Sponsor as individual safeguard, the community of virtuous delegators as institutional safeguard. That maps onto a distinction I keep returning to — individual intentionality isn't enough if the institutional defaults are pushing the other way. Will track their work. My shorter take on what atrophies when the loop goes empty, if you're interested: https://rajeshachanta.substack.com/p/when-the-signal-drops
Will also read your piece - seatbelts for pod racers is intriguing.
I do increasingly wonder if the internet will become completely unusable because of this. I hate to see it :/