Yes, the universities are a problem. But so is the tech industry as it has developed since ChatGPT. It has become a monoculture that's great for AI in the short run but will cripple it in the middle to long range. The industry seems to think that Gary Marus and others like him are Luddite loons who don't know what they're talking about. That's wrong.
And there's a deeper version of his critique in a paper published in 1975 by a forgotten mathematician, Miriam Lipschitz Yevick, the 5th woman to get a PhD in math from MIT back in the late 1940s. In those days there was no place for a woman mathematician in the academy. (When I was an undergraduate at Hopkins in the mid-1960s there were 3 women on a faculty of 500-600.) To the extent that she's remembered it's for her correspondence with the physicist David Bohm, who was in exile in Brazil in the 1950s.
That 1975 paper was entitled "Holographic or fourier logic" and was published in a good journal, Pattern Recognition. I discovered it in a 1978 comment to an article in Brain and Behavioral Science. I made it central to my thinking on the mind and AI, but there was no audience for it back then and so it was forgotten. She proved that in a world with geometrically simple objects (like letter forms and convex polygons) and complex objects (like Chinese ideograms and Rorschach blots) one computing regime is best for dealing with simple objects while another is best for dealing with complex objects. We need both computing regimes working in tandem. The industry has doubled down on one of them.
When ChatGPT appeared in late 2022 I went looking around to see if anyone was citing Yevick. I couldn't find anyone. So I started writing and posting about her work. At that time there was no Wikipedia entry for her. As of January of this year there is. I had nothing to do with it, it came as a surprise, but the article cites my work, so I assume that that's what prompted someone to write the entry. Who knows, maybe in 3 more years someone in the industry will find that article and take it seriously. In any event, the Wikipedia entry is now there for the AIs to train on it. Maybe if someone in the industry comes across her entry through a prompt they'll get curious and check her out. But I'm not holding my breath.
What an interesting and perceptive comment. Have you looked at the similarities and differences between humans and AI in terms of Cognitive Load Theory? The finite computational constraints may be different but they both exist and govern what information gets retained, compressed and ignored. I think this is particular problem in relation to the development of scepticism. It seems to have absorbed the 'people like us' bias from the sources it's been trained on. I've asked it to give links for academic papers to demonstrate points of contention, and on may occasions it has cited work whose methodology or conclusions collapse under even the most basic mathematical scrutiny, or fail on an empirical basis for a host of other reasons. A good example would be swallowing the educational myth that suspensions are in way comparable to classroom disruption in terms of education lost, the latter of which is currently a catastrophic failure in the West, in many instances costing at least two years worth of lost education by the end of secondary education.
Most people don't realise that we are in the capital burn phase (or at least how much of the capital burn is tied up in user consumption) and we may have begun to hit a wall in terms of compute resource use efficiency. At some point AI providers are going to ask users to reach far deeper into their pockets.
Can you give me a reference to cognitive load theory?
The issue of finiteness is an interesting one. I have been saying in various times and places that it seems to me that AI has (implicitly) taken chess as its prototype for AI research. For one thing, we have John McCarthy's well-known article, “Chess as the Drosophila of AI” (1990).
That is, however, a mistake, as I have pointed out in a recent working paper, Computation, Chess, and Language in Artificial Intelligence.* Chess is well-defined, while natural language is not. As a consequence the search space for chess is simple in form, a tree, and well understood. That is not at all the case for natural language. Finally, chess is finite, very large, but finite. That is not at all the case for natural language. Consequently chess is not at all a good paradigm for intelligence in general. Intuitions thus gained from it are likely to be misleading for the general problem.
Wow, that's interesting. Because I've been using AI to edit my writing in a context specific manner, it took my comparison between CLT and AI and used your question to search for papers which are actually exploring my observation about the parallels. It's certainly not the first time I've come across the synchronicity found by people thinking about similar problems around the world (I always find the discovery more than a little edifying), but it doesn't often happen with my AI queries. I shall have to start placing more of my interactions in full into AI. It's leap was entirely unprompted.
I've only read the first few paragraphs of the Introduction to make sure it includes the basic concepts of CLT, but I've already bookmarked it so I can read it with the close scrutiny it seems to deserve at first glance.
I also find the basic graphical representation is a useful primer, so I've included a source with the graphic. You only need the second graphic. Ignore the rest: it's too basic for you:
the degradation of the web and, so too, of search...everyday so exhausting: "Everyone’s learning this the hard way now as the internet is breaking down, search is degrading, nobody can find things they used to, history is disappearing, and the two most reliable places to get an answer are the physical library and AI."
Predictions...hmmm.... It seems to be a match of inertia vs. momentum. I don't see inertia absorbing AI momentum, nor momentum quickly overcoming institutional inertia, so institutional fracture is the most likely near term outcome. Universities will try to achieve punctuated equilibrium, with limited success.
Thank you, Hollis. Your article reminded me of a recent exchange I had with Christopher Thurley about what education is actually for in an age of abundant access to knowledge and increasingly capable agentic AI.
I keep coming back to a Jobs to Be Done question: what are people really hiring universities to do? I’ve written a few articles about this lately.
I experienced part of this tension myself during a master’s program through COVID. The delivery became hybrid almost overnight, but the underlying model barely changed: lectures, credits, thesis, degree. I even suggested unbundling parts of the program for specialists who only needed an update, parents who could not invest years, or simply curious learners who did not need another credential. There was little interest.
That makes me wonder whether universities may understand the technological shift quite well but are still responding rationally to the incentives around them. As long as employers, students, and society continue to reward the credential more reliably than the learning behind it, why would the institution fundamentally change?
And then there is another question I find even harder: what is the job of the professor in this new environment?
Is it to remain the HiEPO in the room—the most highly educated person whose role is to transfer the “right” knowledge—or to become more of a muse who sparks curiosity, challenges assumptions, and helps students discover their own learning trajectory (using the latest technologies) so they continue learning long after the course has ended?
Because the alternative is changing quickly. An increasingly capable agentic AI could serve me a small piece of knowledge, a case, a question, or a challenge every day, personalized to my learning trajectory and delivered exactly when I want it. It could observe where I struggle, adapt the next challenge, and remain available whenever curiosity appears.
So perhaps the scarce contribution of a professor is no longer access to knowledge, but something much harder to automate: helping a student discover what is worth becoming curious about in the first place.
Which leaves me with the question I find hardest: if the degree stopped carrying value tomorrow and an AI could personalize learning around me every day, what would I still hire a university—and especially a professor—to do that would make me choose them over learning along my own trajectory?
Target reached. If life is not about touching other people’s lives, connecting, and exchanging ideas, what else is it about? APIs can exchange information. We can spark something in one another and help each other grow—but only if we overcome our fear of speaking up and stop being silent.
Academia is its own little world, as far as the people getting paid are concerned, it's all about their petty politics and status games; for those paying, directly or indirectly, it has only one function: certification. Not education, not research (though occasionally useful). Certification.
But universities are really not doing their job nearly as well as a few hours of computer-adaptive tests could in 1999 for less than 1/100th the cost, or even the pre-1994 SAT. Years of “education” has a 0.1 predictive validity for job performance, a test of general mental ability 0.65. Adding education to IQ adds 1% to the validity, vs. 20% for adding an integrity test*. The costs to productivity alone are trillions of dollars per year for not using objective measures, without even counting the costs of having the world run by midwits.
For over fifty years, a misreading of the Griggs v. Duke Power Supreme Court opinion has suppressed use of objective tests and made universities gatekeepers of all employment and status. It can't stand much longer, some university with status or firm with good marketing is going to start awarding degrees based on demonstrated ability alone. Then academia's bubble-world will pop, and many living in it will have a difficult time.
*“The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 100 Years of Research Findings” Schmidt & Oh, Working Paper, October 2016, p. 66
Seems like there's a rich theoretical space for higher ed. Like, what's an ontology? who should control them and how? Why did technologists co-opt this term in the first place?
One other thought: perhaps Ed has a greater impact than recognized, especially in light of the NVidia "five layer cake" analogy.
'I am half in love with the typewriter and the telephone. With letters and cables and brief but courteous commands on the telephone to Paris, Berlin, New York, I have fused my many lives into one; I have helped by my assiduity and decision to score those lines on the map there by which the different parts of the world are laced together. I love punctually at ten to come into my room; I love the purple glow of the dark mahogany; I love the table and its sharp edge; and the smooth-running drawers. I love the telephone with its lip stretched to my whisper, and the date on the wall; and the engagement book. Mr Prentice at four; Mr Eyres sharp at four-thirty.
Fascinating and illuminating essay, as always. The university as Bartleby on AI is very thought provoking. Isn't there something admirable about his refusals? In any case, I think Melville is probably the author that somehow best captures this crazy moment.
On the question of how AI will affect institutions, including, but not limited to, education, Samuel Hammond's series of posts are worth reading. First post is here: https://www.secondbest.ca/p/ai-and-leviathan-part-i
Is this why I created a custom GPT for one of my fall courses in the U’s ChatGPT EDU and have never heard back about it being approved or not and am now searching for an option to do the same using a different tool? And why so many of the features are disabled anyway that it’s less capable than a free account? 🤔
AI is not like the telephone! Everyone knew how to use a telephone as soon as it was invented. It replaced a menial task, not a person's intelligence.
You say, "I assumed that a university’s welcome packet to new students and new faculty would say something like this: “you know things that AI does not know, and our goal is to increase both AI and human knowledge.”" But most people do not know things that AI does not know! I interact with AI every day on extremely high-level subjects in mathematics. It is, quite simply, smarter than everyone. Mathematically, there is *absolutely nothing* a typical undergrad student at my institution will be able to do at the end of four years that Claude Opus cannot do. Even Ph.D. students could easily do all of their homework on ChatGPT, and probably even write papers if they wanted to! These days, I'm wondering if there's anything *I" can do that AI won't be able to do in, oh, a year or two? Should we just stop teaching math to make room for real AI-ready skills like asking Claude Code to build something?
You are absolutely correct that AI is the defining change of our time, but if anything you're underestimating it. There is plenty of evidence to suggest that we will all simply be steamrolled by it. Most of us, even us NSF CAREER award winning faculty, have no idea what to do about it. The idea that the university can somehow prepare all its students for what's coming is just laughable. Absolutely no one can prepare anyone for what's coming.
Yes, the universities are a problem. But so is the tech industry as it has developed since ChatGPT. It has become a monoculture that's great for AI in the short run but will cripple it in the middle to long range. The industry seems to think that Gary Marus and others like him are Luddite loons who don't know what they're talking about. That's wrong.
And there's a deeper version of his critique in a paper published in 1975 by a forgotten mathematician, Miriam Lipschitz Yevick, the 5th woman to get a PhD in math from MIT back in the late 1940s. In those days there was no place for a woman mathematician in the academy. (When I was an undergraduate at Hopkins in the mid-1960s there were 3 women on a faculty of 500-600.) To the extent that she's remembered it's for her correspondence with the physicist David Bohm, who was in exile in Brazil in the 1950s.
That 1975 paper was entitled "Holographic or fourier logic" and was published in a good journal, Pattern Recognition. I discovered it in a 1978 comment to an article in Brain and Behavioral Science. I made it central to my thinking on the mind and AI, but there was no audience for it back then and so it was forgotten. She proved that in a world with geometrically simple objects (like letter forms and convex polygons) and complex objects (like Chinese ideograms and Rorschach blots) one computing regime is best for dealing with simple objects while another is best for dealing with complex objects. We need both computing regimes working in tandem. The industry has doubled down on one of them.
When ChatGPT appeared in late 2022 I went looking around to see if anyone was citing Yevick. I couldn't find anyone. So I started writing and posting about her work. At that time there was no Wikipedia entry for her. As of January of this year there is. I had nothing to do with it, it came as a surprise, but the article cites my work, so I assume that that's what prompted someone to write the entry. Who knows, maybe in 3 more years someone in the industry will find that article and take it seriously. In any event, the Wikipedia entry is now there for the AIs to train on it. Maybe if someone in the industry comes across her entry through a prompt they'll get curious and check her out. But I'm not holding my breath.
What an interesting and perceptive comment. Have you looked at the similarities and differences between humans and AI in terms of Cognitive Load Theory? The finite computational constraints may be different but they both exist and govern what information gets retained, compressed and ignored. I think this is particular problem in relation to the development of scepticism. It seems to have absorbed the 'people like us' bias from the sources it's been trained on. I've asked it to give links for academic papers to demonstrate points of contention, and on may occasions it has cited work whose methodology or conclusions collapse under even the most basic mathematical scrutiny, or fail on an empirical basis for a host of other reasons. A good example would be swallowing the educational myth that suspensions are in way comparable to classroom disruption in terms of education lost, the latter of which is currently a catastrophic failure in the West, in many instances costing at least two years worth of lost education by the end of secondary education.
Most people don't realise that we are in the capital burn phase (or at least how much of the capital burn is tied up in user consumption) and we may have begun to hit a wall in terms of compute resource use efficiency. At some point AI providers are going to ask users to reach far deeper into their pockets.
Can you give me a reference to cognitive load theory?
The issue of finiteness is an interesting one. I have been saying in various times and places that it seems to me that AI has (implicitly) taken chess as its prototype for AI research. For one thing, we have John McCarthy's well-known article, “Chess as the Drosophila of AI” (1990).
That is, however, a mistake, as I have pointed out in a recent working paper, Computation, Chess, and Language in Artificial Intelligence.* Chess is well-defined, while natural language is not. As a consequence the search space for chess is simple in form, a tree, and well understood. That is not at all the case for natural language. Finally, chess is finite, very large, but finite. That is not at all the case for natural language. Consequently chess is not at all a good paradigm for intelligence in general. Intuitions thus gained from it are likely to be misleading for the general problem.
*https://www.academia.edu/164885566\Computation_Chess_and_Language_in_Artificial_Intelligence
Wow, that's interesting. Because I've been using AI to edit my writing in a context specific manner, it took my comparison between CLT and AI and used your question to search for papers which are actually exploring my observation about the parallels. It's certainly not the first time I've come across the synchronicity found by people thinking about similar problems around the world (I always find the discovery more than a little edifying), but it doesn't often happen with my AI queries. I shall have to start placing more of my interactions in full into AI. It's leap was entirely unprompted.
I've only read the first few paragraphs of the Introduction to make sure it includes the basic concepts of CLT, but I've already bookmarked it so I can read it with the close scrutiny it seems to deserve at first glance.
I also find the basic graphical representation is a useful primer, so I've included a source with the graphic. You only need the second graphic. Ignore the rest: it's too basic for you:
https://www.barefootteflteacher.com/p/what-is-cognitive-load-theory?hide_intro_popup=true
This second source is the really interesting one. I plan to tackle it with my morning coffee.
https://link.springer.com/article/10.1007/s10462-026-11510-z
the degradation of the web and, so too, of search...everyday so exhausting: "Everyone’s learning this the hard way now as the internet is breaking down, search is degrading, nobody can find things they used to, history is disappearing, and the two most reliable places to get an answer are the physical library and AI."
Very tightly written and a genuine pleasure to read.
Thank you!
silicon vs carbon. silicon is winning. not happy face.
Predictions...hmmm.... It seems to be a match of inertia vs. momentum. I don't see inertia absorbing AI momentum, nor momentum quickly overcoming institutional inertia, so institutional fracture is the most likely near term outcome. Universities will try to achieve punctuated equilibrium, with limited success.
Punctuated equilibrium is right!
Do I win a prize?
This is an excellent piece and a sound warning that universities would be wise to heed. I won’t hold my breath in the hopes they do.
Thank you, Hollis. Your article reminded me of a recent exchange I had with Christopher Thurley about what education is actually for in an age of abundant access to knowledge and increasingly capable agentic AI.
I keep coming back to a Jobs to Be Done question: what are people really hiring universities to do? I’ve written a few articles about this lately.
I experienced part of this tension myself during a master’s program through COVID. The delivery became hybrid almost overnight, but the underlying model barely changed: lectures, credits, thesis, degree. I even suggested unbundling parts of the program for specialists who only needed an update, parents who could not invest years, or simply curious learners who did not need another credential. There was little interest.
That makes me wonder whether universities may understand the technological shift quite well but are still responding rationally to the incentives around them. As long as employers, students, and society continue to reward the credential more reliably than the learning behind it, why would the institution fundamentally change?
And then there is another question I find even harder: what is the job of the professor in this new environment?
Is it to remain the HiEPO in the room—the most highly educated person whose role is to transfer the “right” knowledge—or to become more of a muse who sparks curiosity, challenges assumptions, and helps students discover their own learning trajectory (using the latest technologies) so they continue learning long after the course has ended?
Because the alternative is changing quickly. An increasingly capable agentic AI could serve me a small piece of knowledge, a case, a question, or a challenge every day, personalized to my learning trajectory and delivered exactly when I want it. It could observe where I struggle, adapt the next challenge, and remain available whenever curiosity appears.
So perhaps the scarce contribution of a professor is no longer access to knowledge, but something much harder to automate: helping a student discover what is worth becoming curious about in the first place.
Which leaves me with the question I find hardest: if the degree stopped carrying value tomorrow and an AI could personalize learning around me every day, what would I still hire a university—and especially a professor—to do that would make me choose them over learning along my own trajectory?
This is the conversation I was hoping to spark, yes.
Target reached. If life is not about touching other people’s lives, connecting, and exchanging ideas, what else is it about? APIs can exchange information. We can spark something in one another and help each other grow—but only if we overcome our fear of speaking up and stop being silent.
Academia is its own little world, as far as the people getting paid are concerned, it's all about their petty politics and status games; for those paying, directly or indirectly, it has only one function: certification. Not education, not research (though occasionally useful). Certification.
But universities are really not doing their job nearly as well as a few hours of computer-adaptive tests could in 1999 for less than 1/100th the cost, or even the pre-1994 SAT. Years of “education” has a 0.1 predictive validity for job performance, a test of general mental ability 0.65. Adding education to IQ adds 1% to the validity, vs. 20% for adding an integrity test*. The costs to productivity alone are trillions of dollars per year for not using objective measures, without even counting the costs of having the world run by midwits.
For over fifty years, a misreading of the Griggs v. Duke Power Supreme Court opinion has suppressed use of objective tests and made universities gatekeepers of all employment and status. It can't stand much longer, some university with status or firm with good marketing is going to start awarding degrees based on demonstrated ability alone. Then academia's bubble-world will pop, and many living in it will have a difficult time.
*“The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 100 Years of Research Findings” Schmidt & Oh, Working Paper, October 2016, p. 66
Seems like there's a rich theoretical space for higher ed. Like, what's an ontology? who should control them and how? Why did technologists co-opt this term in the first place?
One other thought: perhaps Ed has a greater impact than recognized, especially in light of the NVidia "five layer cake" analogy.
ZH
This is exactly the conversation I want to have! But presidents are saying let them eat five layer cake on their own time.
Says Louis in The Waves / Virginia Woolf (1931)
'I am half in love with the typewriter and the telephone. With letters and cables and brief but courteous commands on the telephone to Paris, Berlin, New York, I have fused my many lives into one; I have helped by my assiduity and decision to score those lines on the map there by which the different parts of the world are laced together. I love punctually at ten to come into my room; I love the purple glow of the dark mahogany; I love the table and its sharp edge; and the smooth-running drawers. I love the telephone with its lip stretched to my whisper, and the date on the wall; and the engagement book. Mr Prentice at four; Mr Eyres sharp at four-thirty.
My neck aches from nodding agreement at each paragraph.
Some of us are working on the transformation, but as you say, higher ed is resisting. It's in defensive mode.
Fascinating and illuminating essay, as always. The university as Bartleby on AI is very thought provoking. Isn't there something admirable about his refusals? In any case, I think Melville is probably the author that somehow best captures this crazy moment.
Interesting that Preece was so negative about the telephones, because he was quite positive about wireless communications
On the question of how AI will affect institutions, including, but not limited to, education, Samuel Hammond's series of posts are worth reading. First post is here: https://www.secondbest.ca/p/ai-and-leviathan-part-i
Yes those are good pieces.
Is this why I created a custom GPT for one of my fall courses in the U’s ChatGPT EDU and have never heard back about it being approved or not and am now searching for an option to do the same using a different tool? And why so many of the features are disabled anyway that it’s less capable than a free account? 🤔
!!!
!!!
AI is not like the telephone! Everyone knew how to use a telephone as soon as it was invented. It replaced a menial task, not a person's intelligence.
You say, "I assumed that a university’s welcome packet to new students and new faculty would say something like this: “you know things that AI does not know, and our goal is to increase both AI and human knowledge.”" But most people do not know things that AI does not know! I interact with AI every day on extremely high-level subjects in mathematics. It is, quite simply, smarter than everyone. Mathematically, there is *absolutely nothing* a typical undergrad student at my institution will be able to do at the end of four years that Claude Opus cannot do. Even Ph.D. students could easily do all of their homework on ChatGPT, and probably even write papers if they wanted to! These days, I'm wondering if there's anything *I" can do that AI won't be able to do in, oh, a year or two? Should we just stop teaching math to make room for real AI-ready skills like asking Claude Code to build something?
You are absolutely correct that AI is the defining change of our time, but if anything you're underestimating it. There is plenty of evidence to suggest that we will all simply be steamrolled by it. Most of us, even us NSF CAREER award winning faculty, have no idea what to do about it. The idea that the university can somehow prepare all its students for what's coming is just laughable. Absolutely no one can prepare anyone for what's coming.
Oh my, you know your own family history and lore and things that happened in college that AI does not know. These things matter.