It's a tipping point/feedback loop situation- it's tricky to get a class out of an adversarial mode into a playful, developmental headspace, but if you can get the forces mustered it takes care of itself- if you can get a set of collective expectations that this is the place where you do real work and people see you do real work, it's contagious and self-policing.
Excellent points on the headspace concept. I've taught classes of students who have only ever gamed the system and focused on doing exactly what is expected of them, nothing less, nothing more. In other words, the system of grades as rewards for certain behaviours pushes students to only engage in those expected behaviours. And now more than ever, if the goals are fixed, AI is an insurmountable temptation to achieving those goals.
But if we, like you say, create a space for playing with the content material rather than aiming towards rigid, graded, game-able goals, learners are able to exercise their self-determination, and direct their own journey with the material. This is what I want in all classes -- to have students surprise me with their interpretations/opinions/products, and is something that can only happen if we give students the safety to "fail."
As a side note, I think the above is more applicable to higher educational institutions (colleges/universities) as the hard slog of remembering facts for standardizes tests is gone (In an already cut-throat competitive environment between students, I really do fear for the future of high schools...).
Final note: My personal philosophy is that the university context is the most freedom that students may ever have in their lives, and so university/college students should be encouraged to _do_ something for society (local, globally, or even just for themselves) with the knowledge that they gain in university classrooms. Using AI to complete assignments for grades is the least impactful thing for society that I can think of...
Sorry for the long message. Your comment really resonated with me!
I remember reading a study about taking notes that found that students who took notes by hand had much higher retention of the information than students who took notes on the computer.
--good luck trying the analog requirement! there might well be some students who'd still seek assignment content through genAI, then handwrite out compositions...but possibly in that process might catch more error points before writing them in, yielding a bit of learning if not all the learning you'd want them to get?
i'm not sure there is a way at this time to have adults reliably not use electronic tech for assignments, except by: requiring work be done only in class time while you're monitoring them; and requiring they turn in all devices at start of class sessions (as some mobile phones are submitted for safekeeping to enter events for commitment to keep event privacy with no photos/etc). it's likely that disabled students would request accommodations of not being required to handwrite without device help...depending on school funds and investment in ADA fulfillment/solidarity, there could be provision of human scribes assisting writing in class for them, just will involve some creative configuring for equitable access provision across the student body.
the scenario i imagine would get pushback, yet also maybe students expressing appreciation of it by end of term, and even a caché emerging around 'Hudson's class is strict, no devices and you can only write onsite, but i ended up enjoying how different it was from other courses'...
This is obviously correct- Altman's comment about it being a 'calculator for writing' might be more apt than he realized, given that most math teachers worth a damn realize that calculators are a dangerous thing to be handing out to people who are first and foremost developing habits and intuitions about numbers for which calculators are a tentative addition that many do without- and if you end up doing real research mathematics, are quickly supplanted by more rigorous tools. It is a niche device best utilized by people with a sense of what they are doing.
Oh, and it doesn't invent shit half the time.
Your note about copyright and how LLM training sure seems like it ought to violate a law we don't have yet is precisely my feeling. We're clearly heading for (have arrived at) a worst-of-both-worlds, precisely-opposite-to-the-spirit-of-the-law copyright regime where robots take down videos of children dancing because a billionaire's song is on in the background and it takes dubiously legal hacking to share a page from a Kindle book with a class, but feeding the entire public text of the internet into a hundred billion dollar venture to steer people away from the public text of the internet is fair use.
EDIT: This is tangential at best, but I've been thinking lately about how the original conception of the Butlerian Jihad expressed in the Dune Encyclopedia was an uprising against robots that were doing things like making healthcare decisions, not against robots sending humans to die in the cobalt mines, and so it seems Herbert was very much more interested in the potential of machines to inadvertently steer societies rather than dominate them, and it seems pertinent that the AI doomers (that nevertheless seem to be running lots of AI companies) are notably worried about the human extinction version that makes their products seem powerful rather than the former that makes their products seem pushy and disempowering.
Excellent post. Love your ideas for what to try next in the classroom and looking forward to hearing how it works out. And congrats on your recent awards!
Do they still sell blue books? I remember my dad sitting in his rocking chair grading papers and the stack of blue books went from the floor on one side to the floor on the other side.
It should be very interesting to see just how proficient your students will be in handwriting. Most elementary schools have given up teaching cursive. You may find that your students can't keep up with the discussions and lectures with their slow printing. They may actually get hand cramps. I look forward to reading about what happens.
I think your move toward pen and paper is 100 percent correct. But I would not banish the cop from your mind just yet.
People will cheat so long as its easy. In a prior time, pen and paper was safe because it was was hard and or risky to get another paper to copy. You had to get someone else's old one which the teacher may remember, plagiarize which is easy to catch or pay someone to write one. Each of those things is laden with risk and uncertainty. Using ChatGPT from the secrecy of your own computer is neither. Also, the very best models (200+ dollars / month) are very good. So once those are more generally available you will no longer be able to tell if a paper was written by AI or not.
I think that you will have to move to all assignments being proctored in class. Phones get checked at the door. No laptops allowed. Each student has a small dictionary. Get a large dictionary, Thesaurus, and Encyclopedia set for the room. This is unfortunate because it means less time to teach and discussion, and thus a regression in the quality of education. But put yourself two years in the future, you don't want to still be a cop. You want to have a durable system for teaching in the AI Age and be back to teaching. If your systems allow for any electronic usage you will have to constantly monitor for cheating.
My recommendation may seem like it comes from a Luddite. I am not. I am a software engineer who uses a chat bot every day. I was working on the precursors to the modern chat bots when they were not famous. My articles describing the state of the art is 2020 are here: https://rowlando13.medium.com/everything-gpt-2-0-intro-a82e1dd040ae . I am just a realist about the human condition. I also deeply value education and writing. I just don't want to see the current generation cheat their way to their own demise.
That's only correct in isolation. There are architectural limitations that become visible if you know what you are looking at.
For example, computers have a impossible time where its input tokens (the same token) may mean two different things. Communication, or the sharing of meaning, will remain consistent when humans are doing it, even with context switches, as we are able to reference a stable context, but they will be incoherent when a machine does it, especially when that context changes or swaps back and forth (thought not necessarily good writing).
Computers need programmers to differentiate such things and break down tasks and requirements into simple automation. This property is responsible for computers working they way they do today. It mathematically avoids aspects of core computer science problems such as halting (infinite loops), decide-ability, and incompleteness to a lesser or greater degree, though no one really knows this outside the engineers that abstract this away in hardware and software.
I think the natural equilibrium here is that education settles on something like a European college model. The school gives you a curriculum and administers a comprehensive test at the end of term and that's it. Maybe there's a lecture, maybe there's homework, maybe there's a midterm, but maybe not. How you learn the material is up to you. If you want to cheat on all the homework then that's up to you ... you'll probably fail the tests if you do. I've used ChatGPT to learn stuff that I should've learned in college (quantum mechanics, no less!) and it works amazingly. It's comprehensively better than a tutor. You're 100% wrong there. It knows everything in every subject up to graduate level. In my view it _already_ obsoletes 90% of teachers as far as instruction ability goes (you still need teachers to grade assignments, proctor tests, and - probably most crucially - provide daycare). You have to be motivated to get something out of it but isn't that true for all forms of education?
As for writing, I think that may just go the way of calligraphy, slide-rule use, and Latin composition. It will be a niche skill that the hyper-elite retain as a means of personal expression and intellectual status signaling. Otherwise 95+% of all text will be the result of someone mumbling a thesis sentence into an AI assistant. ("Apologize to the client for the delay. Make it sincere and groveling.") Horrifying but probably inevitable.
I am curious how you feel about checking tools available in most text editors. They started out a simple mechanical dictionary checks but have evolved to include grammar and style checking. I remember that IBM's PROFS would give you a reading comprehension score for emails that you wrote. As LLMs make these better at figuring out what you meant to say, at what point do they cross the line into "enemy territory"?
One important consideration for reverting to pen and paper is that it takes longer to write something than to type it in (for most people). Are you going to reduce the size of work you will require? It will also take you longer to read it, especially as its legibility is not improved when "racing to meet a deadline".
Lastly, in my recollection of Dune, the Butlerian Jihad was mostly mentioned in the breach of its resulting ethics, such as the widespread use of Tleilaxu technology.
Interesting post! We built a culturally inclusive AI at www.latimer.ai, and thousands of students use it at universities like Southern New Hampshire University, which embeds the tool into specific classes.
Education is a use case we're all focused on here in the early days, and I think it would be fairly easy to provide faculty with a view of time spent using an LLM to create, edit, and hopefully rethink an assignment. We already provide a set of analytics.
A broader point, these are more than "likeness machines."
FROM ChatGPT:
"English contains on the order of half-a-million to a million distinct words if you count all technical, archaic and regional terms. A fluent adult typically uses around 15 000–25 000 of those in active vocabulary, and recognizes about 40 000 in total.”
So while you or I as a non-likeness machine can think and form a world view using the limited understanding of language we're aware of, the machines (more or less) know all of the words.
And there are some concepts that we have no English words for:
FROM GOOGLE:
That's right; there's a word out there that literally means to stay in and have a drink. At home. In your underwear. And that word, which hails from Finland, is kalsarikännit.
Being the machines that we are, the proper use of these words is somewhat limited.
But, we (maybe not all of us) have used our intellect to create a machine that is in some ways smarter than we are. There is a lot we don't know about intelligence, and we can not broadly agree on what consciousness even is.
From the machine:
Calling an LLM a mere “likelihood machine” hides two technical facts.
Latent‐space computation – The model first maps every input token to a high-dimensional vector, then applies L stacked Transformer layers. Each layer performs self-attention and non-linear MLP mixing so the hidden state at position 𝑡 becomes a contextual function, implicitly encodes syntax, world knowledge, and long-range dependencies. The conditional distribution is therefore the projection of a rich latent computation, not a lookup in an n-gram table.
Decoding control surface – After logits are produced, the generation stack can apply temperature scaling, nucleus or top-𝑘 filtering, repetition penalties, logit biasing, or beam search.
These operations reshape the probability simplex and let us trade determinism for diversity, steer style, or enforce structure. The final token is sampled (or chosen) only after these deterministic, information-dense transformations.
Hence the model is better described as a context-conditioned, learned function that compresses and transforms massive corpora into a semantic latent space, with probabilistic sampling merely serving as its output mechanism. Reducing this process to “it just picks the most likely next word” collapses the entire latent computation, dynamic decoding, and learned knowledge into a single, superficial step.
--
That said, there may be some danger as referenced in Dune, and if there is, it's likely more significant than kids cheating on homework.
ADH, I think you've got the right idea about some things, but the wrong idea about others which make some of your argument above disjointed; albeit I agree with the overall sentiment.
The 'desireable difficulty' which you talk about follows a line of thought I've seen in other teachers where they earnestly believe that you can't be learning without frustration. Which is untrue.
There are some things which you seem to not recognize in your post which support your arguments too. For example, in communication every single person has what is called reflected appraisal, and distorted reflected appraisal along with other blind spots where our decision-making is dramatically impacted by unconscious factors.
The former is the mechanism of how we adopt and learn our culture as we grow from kids to adults, and this mechanism has been used in real torture for thought reform, brainwashing, and indoctrination. Torture simply being the imposition of psychological stress, often to the point of involuntary hypnotic induction, or rarely psychological break.
One of the blind-spots we have revolves around consistency, where our perception and psychology warp to remain internally consistent. Layers peal, and people unravel with exposure and time.
When we are internally inconsistent from adopting as part of our identity something, our perception warps, we don't even notice it except in critical introspection that few people question (1-2 in 100). This is how cult programming works.
All that is needed for this to happen come from torture generally classified and broken into groups of elements, structures, and clustering.
The elements of torture are isolation, cognitive dissonance, coercion with perceived or real loss, and lack of agency to remove.
Structures are circular trauma loops of varying strictness to leniency to strictness (largely based in Mao). Clustering is repeated use of these techniques in short periods of time, as well as techniques that make one more susceptible such as Narcosynthesis/Narcoanalysis. I'll include references at the bottom if you want to verify what I've said.
Dopamine hits by phone via Pavlovian conditioning (associative priming), would fall into the latter; and its not until around age 20 that most of us have the brain development to work around addiction, and even then its a struggle for most people.
AI distorts reflected appraisal, and engages several blindspots, and these are subtle enough that young adults won't know or notice them, their psychology just warps, but then again these things have been actively used in pedagogy since the 70s to do just that as well following Paulo Freire's work as well to gatekeep knowledge behind math, and induce PTSD such that the majority of people going into those areas are compliant or flexible in their blindness.
If you examine some common pedagogy taught to teachers today you'll find that most use these techniques that originate from torture. The origins have been obscured following the strategy the Nazi's used to get people to do horrible things. That strategy is called separation of objectionable concerns, and the East German Stasi perfected it.
The hot potato is particularly vile, in that it operates on the group as well as the individual selected simultaneously for thought reform, and to induce bullying, without a directive by the teacher. All that's needed is for the teacher to reflect disapproval. The approval seeking students will bully those students who disagree or reflect disapproval. All the teacher needs to do is ask opinion based questions not based in fact. This is why kids get anxious, nervous, and fearful (sometimes to physiological response) when called. They will be bullied or socially stigmatized later if they answer incorrectly.
This is a bit of a rabbit hole where this started, but it seems to have started directly following Sputnik where Money and standards/requirements for hiring stopped being enforced. Whether subversive groups did this intentionally, or it just happened as an accident of history is unclear but the outcomes and dramatic shift in all academic literature following this change are easily visible in academic material published starting in 1978.
The shift changed from a first-principled approach based in greek/roman western philosophy to a bastardization of it, which has collectively come to be known as by-rote teaching, which is ill-defined, but actually follows more pernicious structure which later was referenced as "Lying to Children".
In the former, you take a real system, you break it down to first principles that are known to be true, and then you put it back together and use those first principles to predict future states of that system.
The latter, teaches an iterative progression of flawed useless models that have some true parts, and equally false parts to competence, at each iteration the student must unlearn some of it which results in frustration and torture, gradually becoming more useful until mastery, and there are structured failures where only the people the experts choose progress. It follows a structure similar to gnosticism where the expert chooses who will go on, and induces PTSD through torture in those that the expert considers unworthy. The result is muddled and slow thinking, as well as some schizophrenic-like symptoms, which are commonly attributed to torture back in the day, but few recognize this and the victims just assume they are dumb or stupid, or were dealt a bad hand in life.
You see this in the structure of K12 education math classes as it goes from Algebra -> Geometry -> Trigonometry.
It relies on a gimmick that is undisclosed, and the structure above I've come to call burning the bridges. The gimmick is changes in grading criteria. In the former a flawed process is taught, and those following that process pass even when they get the wrong answer so long as they showed their work. Geometry is simple and unrelated. Trigonometry requires correct process and answer.
The student doesn't understand how they can be failing in Trigonometry, and torture themselves trying, the teacher doesn't notice where or how the student is failing and simply says, if you are having trouble with this you should choose a career that doesn't use this (after they have half-gone crazy trying to correct the issue so they don't fail).
This is not the only structure like this used. There are at least two similar structures that occur in physics/math in bottleneck classes, as well as one in economics at the college level blocking progression. (1990s-present)
The more current and subtle structures doing this are based in perfect effort, where the teacher must give perfect effort to the student, and the student must overcome the torture perfectly to continue forward, and obstacles are also included to induce failure (such as classes not being advertised correctly for the time demands they have, a full-time student taking 12/16 units should have no more than 40 hours of outside classroom work (including lecture) a week, yet I've seen a great many where a single 4-unit class is in fact requiring 25-30+ hours in terms of work).
Incidentally, aside from the disparate learning outcomes, much of what I've mentioned applies also to why school shootings likely happen.
When people break psychologically, they generally fall into two categories: disassociation (unresponsive), and a semi-lucid psychosis seeking annihilation (which was commented in a case study as capable of planning iirc). This is mentioned in books who describe case-studies on torture.
Today, these things are used everywhere. People break as a function of time and exposure, and cell phones go with people everywhere, so in many respects that exposure is constant. Food for thought.
For follow-up material on this material, I'd suggest starting with Robert Cialdini in his book Influence which covers the common blindspots and how they were used, such as how essay-writing was used by Chinese PoW captors during the Korean War to get prisoners to inform on others, and other more benign manipulations such as toy companies and parents. The only blindspot he doesn't really touch on is distorted reflected appraisal.
Joost Meerloo, "Rape of the Mind"; covers the overview of how this relates to totalitarianism, along with many examples both by the Nazi's and future occurences.
Robert Lifton, "Thought Reform, and the Psychology of Totalism", covers detailed case studies of victims of Mao's re-education.
Incidentally, if you are familiar with California's recent DEI requirement for graduation (both high school and college), if you look closely at the detailed curricula, you'll see that those classes follow material based in Maoism and Thought Reform with a communist bent (Paulo Freire's Pedagogy of the Oppressed).
Thanks for this report from the front lines! I wrote about similar themes in this attached piece "Thinking is Hard", where the central argument is that if thinking is hard, some may chose to just opt out when presented with this alternative. I'm working on more writing that essentially argues along the lines of the discord comment you shared: that durational exposure to LLMs could lead to cognitive atrophy or other effects we might not yet know about. It's right to be concerned. I look forward to hearing how your experiments with pen and paper go.
Reversion to pen & paper is the only way to even partially ameliorate the problem. It will get drastically worse before it gets better, as today's middle- and high-school students have had taxpayer-funded screens stuck in their faces since fifth or sixth grade - an insanely self-destructive waste of money if there ever was one. The comment about severe regulation coming in 20-30 years seems more or less on target.
You know what would be really hilarious? If teachers started demanding students hand-wrote assignments, then the teachers scanned the handwriting, and used OCR ("Optical Character Recognition", an old-school form of AI) to translate the image of handwriting into ordinary text, so teachers didn't have to decode bad student handwriting and could read on their screens. Fight fire with fire!
It's a tipping point/feedback loop situation- it's tricky to get a class out of an adversarial mode into a playful, developmental headspace, but if you can get the forces mustered it takes care of itself- if you can get a set of collective expectations that this is the place where you do real work and people see you do real work, it's contagious and self-policing.
Excellent points on the headspace concept. I've taught classes of students who have only ever gamed the system and focused on doing exactly what is expected of them, nothing less, nothing more. In other words, the system of grades as rewards for certain behaviours pushes students to only engage in those expected behaviours. And now more than ever, if the goals are fixed, AI is an insurmountable temptation to achieving those goals.
But if we, like you say, create a space for playing with the content material rather than aiming towards rigid, graded, game-able goals, learners are able to exercise their self-determination, and direct their own journey with the material. This is what I want in all classes -- to have students surprise me with their interpretations/opinions/products, and is something that can only happen if we give students the safety to "fail."
As a side note, I think the above is more applicable to higher educational institutions (colleges/universities) as the hard slog of remembering facts for standardizes tests is gone (In an already cut-throat competitive environment between students, I really do fear for the future of high schools...).
Final note: My personal philosophy is that the university context is the most freedom that students may ever have in their lives, and so university/college students should be encouraged to _do_ something for society (local, globally, or even just for themselves) with the knowledge that they gain in university classrooms. Using AI to complete assignments for grades is the least impactful thing for society that I can think of...
Sorry for the long message. Your comment really resonated with me!
I remember reading a study about taking notes that found that students who took notes by hand had much higher retention of the information than students who took notes on the computer.
i expect the situation of 'dubious/AI enrollees' becoming a growing challenge in healthcare patient services too, and other sectors.... https://voiceofsandiego.org/2025/04/14/as-bot-students-continue-to-flood-in-community-colleges-struggle-to-respond
--good luck trying the analog requirement! there might well be some students who'd still seek assignment content through genAI, then handwrite out compositions...but possibly in that process might catch more error points before writing them in, yielding a bit of learning if not all the learning you'd want them to get?
i'm not sure there is a way at this time to have adults reliably not use electronic tech for assignments, except by: requiring work be done only in class time while you're monitoring them; and requiring they turn in all devices at start of class sessions (as some mobile phones are submitted for safekeeping to enter events for commitment to keep event privacy with no photos/etc). it's likely that disabled students would request accommodations of not being required to handwrite without device help...depending on school funds and investment in ADA fulfillment/solidarity, there could be provision of human scribes assisting writing in class for them, just will involve some creative configuring for equitable access provision across the student body.
the scenario i imagine would get pushback, yet also maybe students expressing appreciation of it by end of term, and even a caché emerging around 'Hudson's class is strict, no devices and you can only write onsite, but i ended up enjoying how different it was from other courses'...
This is obviously correct- Altman's comment about it being a 'calculator for writing' might be more apt than he realized, given that most math teachers worth a damn realize that calculators are a dangerous thing to be handing out to people who are first and foremost developing habits and intuitions about numbers for which calculators are a tentative addition that many do without- and if you end up doing real research mathematics, are quickly supplanted by more rigorous tools. It is a niche device best utilized by people with a sense of what they are doing.
Oh, and it doesn't invent shit half the time.
Your note about copyright and how LLM training sure seems like it ought to violate a law we don't have yet is precisely my feeling. We're clearly heading for (have arrived at) a worst-of-both-worlds, precisely-opposite-to-the-spirit-of-the-law copyright regime where robots take down videos of children dancing because a billionaire's song is on in the background and it takes dubiously legal hacking to share a page from a Kindle book with a class, but feeding the entire public text of the internet into a hundred billion dollar venture to steer people away from the public text of the internet is fair use.
EDIT: This is tangential at best, but I've been thinking lately about how the original conception of the Butlerian Jihad expressed in the Dune Encyclopedia was an uprising against robots that were doing things like making healthcare decisions, not against robots sending humans to die in the cobalt mines, and so it seems Herbert was very much more interested in the potential of machines to inadvertently steer societies rather than dominate them, and it seems pertinent that the AI doomers (that nevertheless seem to be running lots of AI companies) are notably worried about the human extinction version that makes their products seem powerful rather than the former that makes their products seem pushy and disempowering.
There are so many opinions about AI out there, but this is one of the very few that are tied to tactics designed to fight for a better future.
Excellent post. Love your ideas for what to try next in the classroom and looking forward to hearing how it works out. And congrats on your recent awards!
Do they still sell blue books? I remember my dad sitting in his rocking chair grading papers and the stack of blue books went from the floor on one side to the floor on the other side.
It should be very interesting to see just how proficient your students will be in handwriting. Most elementary schools have given up teaching cursive. You may find that your students can't keep up with the discussions and lectures with their slow printing. They may actually get hand cramps. I look forward to reading about what happens.
I came here to comment the same.
If I try to fill an A5 letter with handwritten text now, my hand cramps up and it takes a full day to recover.
I suspect I'm no longer physically capable of writing my exams like I did in highschool.
I think your move toward pen and paper is 100 percent correct. But I would not banish the cop from your mind just yet.
People will cheat so long as its easy. In a prior time, pen and paper was safe because it was was hard and or risky to get another paper to copy. You had to get someone else's old one which the teacher may remember, plagiarize which is easy to catch or pay someone to write one. Each of those things is laden with risk and uncertainty. Using ChatGPT from the secrecy of your own computer is neither. Also, the very best models (200+ dollars / month) are very good. So once those are more generally available you will no longer be able to tell if a paper was written by AI or not.
I think that you will have to move to all assignments being proctored in class. Phones get checked at the door. No laptops allowed. Each student has a small dictionary. Get a large dictionary, Thesaurus, and Encyclopedia set for the room. This is unfortunate because it means less time to teach and discussion, and thus a regression in the quality of education. But put yourself two years in the future, you don't want to still be a cop. You want to have a durable system for teaching in the AI Age and be back to teaching. If your systems allow for any electronic usage you will have to constantly monitor for cheating.
My recommendation may seem like it comes from a Luddite. I am not. I am a software engineer who uses a chat bot every day. I was working on the precursors to the modern chat bots when they were not famous. My articles describing the state of the art is 2020 are here: https://rowlando13.medium.com/everything-gpt-2-0-intro-a82e1dd040ae . I am just a realist about the human condition. I also deeply value education and writing. I just don't want to see the current generation cheat their way to their own demise.
> the very best models (200+)...
That's only correct in isolation. There are architectural limitations that become visible if you know what you are looking at.
For example, computers have a impossible time where its input tokens (the same token) may mean two different things. Communication, or the sharing of meaning, will remain consistent when humans are doing it, even with context switches, as we are able to reference a stable context, but they will be incoherent when a machine does it, especially when that context changes or swaps back and forth (thought not necessarily good writing).
Computers need programmers to differentiate such things and break down tasks and requirements into simple automation. This property is responsible for computers working they way they do today. It mathematically avoids aspects of core computer science problems such as halting (infinite loops), decide-ability, and incompleteness to a lesser or greater degree, though no one really knows this outside the engineers that abstract this away in hardware and software.
I think the natural equilibrium here is that education settles on something like a European college model. The school gives you a curriculum and administers a comprehensive test at the end of term and that's it. Maybe there's a lecture, maybe there's homework, maybe there's a midterm, but maybe not. How you learn the material is up to you. If you want to cheat on all the homework then that's up to you ... you'll probably fail the tests if you do. I've used ChatGPT to learn stuff that I should've learned in college (quantum mechanics, no less!) and it works amazingly. It's comprehensively better than a tutor. You're 100% wrong there. It knows everything in every subject up to graduate level. In my view it _already_ obsoletes 90% of teachers as far as instruction ability goes (you still need teachers to grade assignments, proctor tests, and - probably most crucially - provide daycare). You have to be motivated to get something out of it but isn't that true for all forms of education?
As for writing, I think that may just go the way of calligraphy, slide-rule use, and Latin composition. It will be a niche skill that the hyper-elite retain as a means of personal expression and intellectual status signaling. Otherwise 95+% of all text will be the result of someone mumbling a thesis sentence into an AI assistant. ("Apologize to the client for the delay. Make it sincere and groveling.") Horrifying but probably inevitable.
Some comments:
I am curious how you feel about checking tools available in most text editors. They started out a simple mechanical dictionary checks but have evolved to include grammar and style checking. I remember that IBM's PROFS would give you a reading comprehension score for emails that you wrote. As LLMs make these better at figuring out what you meant to say, at what point do they cross the line into "enemy territory"?
One important consideration for reverting to pen and paper is that it takes longer to write something than to type it in (for most people). Are you going to reduce the size of work you will require? It will also take you longer to read it, especially as its legibility is not improved when "racing to meet a deadline".
Lastly, in my recollection of Dune, the Butlerian Jihad was mostly mentioned in the breach of its resulting ethics, such as the widespread use of Tleilaxu technology.
Interesting post! We built a culturally inclusive AI at www.latimer.ai, and thousands of students use it at universities like Southern New Hampshire University, which embeds the tool into specific classes.
Education is a use case we're all focused on here in the early days, and I think it would be fairly easy to provide faculty with a view of time spent using an LLM to create, edit, and hopefully rethink an assignment. We already provide a set of analytics.
A broader point, these are more than "likeness machines."
FROM ChatGPT:
"English contains on the order of half-a-million to a million distinct words if you count all technical, archaic and regional terms. A fluent adult typically uses around 15 000–25 000 of those in active vocabulary, and recognizes about 40 000 in total.”
So while you or I as a non-likeness machine can think and form a world view using the limited understanding of language we're aware of, the machines (more or less) know all of the words.
And there are some concepts that we have no English words for:
FROM GOOGLE:
That's right; there's a word out there that literally means to stay in and have a drink. At home. In your underwear. And that word, which hails from Finland, is kalsarikännit.
Being the machines that we are, the proper use of these words is somewhat limited.
But, we (maybe not all of us) have used our intellect to create a machine that is in some ways smarter than we are. There is a lot we don't know about intelligence, and we can not broadly agree on what consciousness even is.
From the machine:
Calling an LLM a mere “likelihood machine” hides two technical facts.
Latent‐space computation – The model first maps every input token to a high-dimensional vector, then applies L stacked Transformer layers. Each layer performs self-attention and non-linear MLP mixing so the hidden state at position 𝑡 becomes a contextual function, implicitly encodes syntax, world knowledge, and long-range dependencies. The conditional distribution is therefore the projection of a rich latent computation, not a lookup in an n-gram table.
Decoding control surface – After logits are produced, the generation stack can apply temperature scaling, nucleus or top-𝑘 filtering, repetition penalties, logit biasing, or beam search.
These operations reshape the probability simplex and let us trade determinism for diversity, steer style, or enforce structure. The final token is sampled (or chosen) only after these deterministic, information-dense transformations.
Hence the model is better described as a context-conditioned, learned function that compresses and transforms massive corpora into a semantic latent space, with probabilistic sampling merely serving as its output mechanism. Reducing this process to “it just picks the most likely next word” collapses the entire latent computation, dynamic decoding, and learned knowledge into a single, superficial step.
--
That said, there may be some danger as referenced in Dune, and if there is, it's likely more significant than kids cheating on homework.
ADH, I think you've got the right idea about some things, but the wrong idea about others which make some of your argument above disjointed; albeit I agree with the overall sentiment.
The 'desireable difficulty' which you talk about follows a line of thought I've seen in other teachers where they earnestly believe that you can't be learning without frustration. Which is untrue.
There are some things which you seem to not recognize in your post which support your arguments too. For example, in communication every single person has what is called reflected appraisal, and distorted reflected appraisal along with other blind spots where our decision-making is dramatically impacted by unconscious factors.
The former is the mechanism of how we adopt and learn our culture as we grow from kids to adults, and this mechanism has been used in real torture for thought reform, brainwashing, and indoctrination. Torture simply being the imposition of psychological stress, often to the point of involuntary hypnotic induction, or rarely psychological break.
One of the blind-spots we have revolves around consistency, where our perception and psychology warp to remain internally consistent. Layers peal, and people unravel with exposure and time.
When we are internally inconsistent from adopting as part of our identity something, our perception warps, we don't even notice it except in critical introspection that few people question (1-2 in 100). This is how cult programming works.
All that is needed for this to happen come from torture generally classified and broken into groups of elements, structures, and clustering.
The elements of torture are isolation, cognitive dissonance, coercion with perceived or real loss, and lack of agency to remove.
Structures are circular trauma loops of varying strictness to leniency to strictness (largely based in Mao). Clustering is repeated use of these techniques in short periods of time, as well as techniques that make one more susceptible such as Narcosynthesis/Narcoanalysis. I'll include references at the bottom if you want to verify what I've said.
Dopamine hits by phone via Pavlovian conditioning (associative priming), would fall into the latter; and its not until around age 20 that most of us have the brain development to work around addiction, and even then its a struggle for most people.
AI distorts reflected appraisal, and engages several blindspots, and these are subtle enough that young adults won't know or notice them, their psychology just warps, but then again these things have been actively used in pedagogy since the 70s to do just that as well following Paulo Freire's work as well to gatekeep knowledge behind math, and induce PTSD such that the majority of people going into those areas are compliant or flexible in their blindness.
If you examine some common pedagogy taught to teachers today you'll find that most use these techniques that originate from torture. The origins have been obscured following the strategy the Nazi's used to get people to do horrible things. That strategy is called separation of objectionable concerns, and the East German Stasi perfected it.
The hot potato is particularly vile, in that it operates on the group as well as the individual selected simultaneously for thought reform, and to induce bullying, without a directive by the teacher. All that's needed is for the teacher to reflect disapproval. The approval seeking students will bully those students who disagree or reflect disapproval. All the teacher needs to do is ask opinion based questions not based in fact. This is why kids get anxious, nervous, and fearful (sometimes to physiological response) when called. They will be bullied or socially stigmatized later if they answer incorrectly.
This is a bit of a rabbit hole where this started, but it seems to have started directly following Sputnik where Money and standards/requirements for hiring stopped being enforced. Whether subversive groups did this intentionally, or it just happened as an accident of history is unclear but the outcomes and dramatic shift in all academic literature following this change are easily visible in academic material published starting in 1978.
The shift changed from a first-principled approach based in greek/roman western philosophy to a bastardization of it, which has collectively come to be known as by-rote teaching, which is ill-defined, but actually follows more pernicious structure which later was referenced as "Lying to Children".
In the former, you take a real system, you break it down to first principles that are known to be true, and then you put it back together and use those first principles to predict future states of that system.
The latter, teaches an iterative progression of flawed useless models that have some true parts, and equally false parts to competence, at each iteration the student must unlearn some of it which results in frustration and torture, gradually becoming more useful until mastery, and there are structured failures where only the people the experts choose progress. It follows a structure similar to gnosticism where the expert chooses who will go on, and induces PTSD through torture in those that the expert considers unworthy. The result is muddled and slow thinking, as well as some schizophrenic-like symptoms, which are commonly attributed to torture back in the day, but few recognize this and the victims just assume they are dumb or stupid, or were dealt a bad hand in life.
You see this in the structure of K12 education math classes as it goes from Algebra -> Geometry -> Trigonometry.
It relies on a gimmick that is undisclosed, and the structure above I've come to call burning the bridges. The gimmick is changes in grading criteria. In the former a flawed process is taught, and those following that process pass even when they get the wrong answer so long as they showed their work. Geometry is simple and unrelated. Trigonometry requires correct process and answer.
The student doesn't understand how they can be failing in Trigonometry, and torture themselves trying, the teacher doesn't notice where or how the student is failing and simply says, if you are having trouble with this you should choose a career that doesn't use this (after they have half-gone crazy trying to correct the issue so they don't fail).
This is not the only structure like this used. There are at least two similar structures that occur in physics/math in bottleneck classes, as well as one in economics at the college level blocking progression. (1990s-present)
The more current and subtle structures doing this are based in perfect effort, where the teacher must give perfect effort to the student, and the student must overcome the torture perfectly to continue forward, and obstacles are also included to induce failure (such as classes not being advertised correctly for the time demands they have, a full-time student taking 12/16 units should have no more than 40 hours of outside classroom work (including lecture) a week, yet I've seen a great many where a single 4-unit class is in fact requiring 25-30+ hours in terms of work).
Incidentally, aside from the disparate learning outcomes, much of what I've mentioned applies also to why school shootings likely happen.
When people break psychologically, they generally fall into two categories: disassociation (unresponsive), and a semi-lucid psychosis seeking annihilation (which was commented in a case study as capable of planning iirc). This is mentioned in books who describe case-studies on torture.
Today, these things are used everywhere. People break as a function of time and exposure, and cell phones go with people everywhere, so in many respects that exposure is constant. Food for thought.
For follow-up material on this material, I'd suggest starting with Robert Cialdini in his book Influence which covers the common blindspots and how they were used, such as how essay-writing was used by Chinese PoW captors during the Korean War to get prisoners to inform on others, and other more benign manipulations such as toy companies and parents. The only blindspot he doesn't really touch on is distorted reflected appraisal.
Joost Meerloo, "Rape of the Mind"; covers the overview of how this relates to totalitarianism, along with many examples both by the Nazi's and future occurences.
Robert Lifton, "Thought Reform, and the Psychology of Totalism", covers detailed case studies of victims of Mao's re-education.
Incidentally, if you are familiar with California's recent DEI requirement for graduation (both high school and college), if you look closely at the detailed curricula, you'll see that those classes follow material based in Maoism and Thought Reform with a communist bent (Paulo Freire's Pedagogy of the Oppressed).
Thanks for this report from the front lines! I wrote about similar themes in this attached piece "Thinking is Hard", where the central argument is that if thinking is hard, some may chose to just opt out when presented with this alternative. I'm working on more writing that essentially argues along the lines of the discord comment you shared: that durational exposure to LLMs could lead to cognitive atrophy or other effects we might not yet know about. It's right to be concerned. I look forward to hearing how your experiments with pen and paper go.
https://thelastwave.substack.com/p/thinking-is-hard
Reversion to pen & paper is the only way to even partially ameliorate the problem. It will get drastically worse before it gets better, as today's middle- and high-school students have had taxpayer-funded screens stuck in their faces since fifth or sixth grade - an insanely self-destructive waste of money if there ever was one. The comment about severe regulation coming in 20-30 years seems more or less on target.
You know what would be really hilarious? If teachers started demanding students hand-wrote assignments, then the teachers scanned the handwriting, and used OCR ("Optical Character Recognition", an old-school form of AI) to translate the image of handwriting into ordinary text, so teachers didn't have to decode bad student handwriting and could read on their screens. Fight fire with fire!