Teaching Humanities in the AI Era: Is this the Bargaining or the Acceptance Stage?
“What, again, is education? The non-coercive rearranging of desire.”
That’s D. Graham Burnett, professor of history at Princeton University, in an article in The New Yorker that you should read.
In the article, Burnett, whose area is history of science and who has focused quite a bit on the topic of attention, documents some of his experiments using large language models (LLMs) as research and teaching tools. In this post I’ll focus on the teaching side.
In one of his courses, “Attention and Modernity: Mind, Media, and the Senses”, he gave the students an assignment that had them make use of one of the LLMs.
An assignment in my class asked students to engage one of the new A.I. tools in a conversation about the history of attention. The idea was to let them take a topic that they now understood in some depth and explore what these systems could do with it. It was also a chance to confront the attention economy’s “killer app”: totally algorithmic pseudo-persons who are sensitive, competent, and infinitely patient; know everything about everyone; and will, of course, be turned to the business of extracting money from us. These systems promise a new mode of attention capture—what some are calling the “intimacy economy” (“human fracking” comes closer to the truth). The assignment was simple: have a conversation with a chatbot about the history of attention, edit the text down to four pages, and turn it in.
How did the students do? Burnett says:
Reading the results, on my living-room couch, turned out to be the most profound experience of my teaching career. I’m not sure how to describe it. In a basic way, I felt I was watching a new kind of creature being born, and also watching a generation come face to face with that birth: an encounter with something part sibling, part rival, part careless child-god, part mechanomorphic shadow—an alien familiar.
He then proceeds to provide several examples of the truly fascinating work his students submitted. Seriously, read the article.
What’s his takeaway from this “for those who are responsible for the humanistic tradition—those of us who serve as custodians of historical consciousness”? Those whose job is “helping others hold those artifacts and insights in their hands, however briefly, and of considering what ought to be reserved from the ever-sucking vortex of oblivion—and why”?
Adopting Gayatri Chakravorty Spivak’s definition of education as the “non-coercive rearranging of desire,” he sees an opportunity, and seems to accept that humanities education from here on in has to be different:
You can no longer make students do the reading or the writing. So what’s left? Only this: give them work they want to do. And help them want to do it. What, again, is education? The non-coercive rearranging of desire.
Oh… just that? Just give them work they want to do, and help them want to do it?
Um, hey, philosophers—have any of you tried doing this?
What’s that? Only every semester since you started teaching?
So you’re good then?
No?
Huh.
Burnett’s sharing of his experiences teaching with AI is indeed interesting and helpful. But, right off the bat, we should be cautious about generalizing from them. After all, he is teaching Princeton students… at Princeton. In terms of intelligence, knowledge, motivation, environment, they are not representative of college students in the United States. That they are especially susceptible to help that gets them to want to do work their professor thinks is worthwhile—that there is academic work they want to do—does not tell us that much about the over 99% of US undergraduates not enrolled in Ivy League or otherwise elite institutions.
Burnett’s course seems to have had a rather small enrollment, which is necessary for the opportunities for (and expectations of) in-person discussion of the students’ projects at a level sufficient to help incentivize hard work and deter fakery. And a light teaching load no doubt helps, too. What about lower level courses with enrollments in the 40- to 240-person range? And what about instructors teaching two or three or more other courses simultaneously? How common is it for instructors to have the time inside or outside of class to run such an assignment well?
I think that many academics who “give them work they want to do and help them want to do it” will be thinking, “yes, but…” What’s the but?
- …but students have all of the world’s entertainment and all of their friends with them at all times via devices that constantly remind them of this fact. In the flood of real and apparent fun, it is ever harder for professors to get a firm enough hold on their students’ attention to help them want work.
- …but many students seem to be in college merely for the degree, so they can then go off and get a job that degree qualifies them for. I suspect many were never adequately made aware that learning might have anything beyond the crudest instrumental value.
- …but students are busy with various activities and many have jobs; even if they wanted to spend more time on schoolwork it’s not difficult to see how it would seem more rational to use the AI tools at their disposal because they’re more “efficient”.
Still, maybe this is too much “easier said than done” and not enough creative brainstorming about the various ways to help various kinds of students at various kinds of institutions want to do some version of the work we think they should do. Your suggestions, as always, are welcome.
Burnett asks, about AI tools, “Do they herald the end of ‘the humanities’?” He answers: “In one sense, absolutely.” You know a “but” is coming, and indeed it does come, followed by the quote which began this post. Pessimistically, the meaning of “one sense” is “for most of you.” It would be nice to not be so pessimistic. Can you help?
I note that third stage of grief, “bargaining,” according to Wikipedia, “involves the hope that the individual can avoid a cause of grief. Usually, the negotiation for an extended life is made in exchange for a reformed lifestyle.”
I think I might have a useful perspective here. I teach at a public system campus, have the equivalent of a 3:4 load, and when I teach ethics, it’s multiple sections of 35 students. In other words, I do not teach at Princeton.
There are two things that stick out to me:
First, Weinberg’s point about helping students do work they want to do. He frames it skeptically, which I think is understandable to a certain degree.
But—and I cannot overstate this—learning research suggests this is genuinely the best way to get students to learn and learn deeply. It’s called “situated skill development” and is basically getting students to apply their learning via a project they find meaningful. Yes, this is hard work, especially in a discipline students aren’t familiar with and probably don’t enter into with a sense that it will be useful beyond ticking some distribution box. My campus is entirely health science-focused, so this is literally every student I have. But if you can find a way to do this, the results are EXPLOSIVE.
Weinberg goes on:
I don’t know how common it is, but I feel like I’ve done it. My ethics class revolves around an Apology assignment where students synthesize what they’ve learned from class to create their own view of a flourishing life. In my other classes, I work with students to understand what they want to say about what they’ve learned in class and how they want to say it. (My class on the ethics and politics of abortion, for example, is currently putting together a gallery display for the university research symposium as a way of engaging with feminist philosophical insights into different eras of abortion in the US.) Yes, it creates a lot of work on my end, but the results are worth it. Doing things this way seems to radically disincentivize using tools like AI because students recognize they want to do the work themselves and say something genuine. (One other thing I do that seems to help in this regard is alternative grading like contract grading. It seems to get students to focus on genuine learning in a way that traditional grading doesn’t. I haven’t done traditional grading in almost five years and you’ll kill me before you get me to go back.)
You can see some examples of past work here: https://bsky.app/profile/bcnjake.bsky.social/post/3kaaadfi3gg2y
Again, I teach the equivalent of a 4:3 with two and sometimes as many as four preps. It isn’t the heaviest load in the world but also isn’t a 2:2 with one prep, TAs to do the grading and 1/3 of the teaching, and a ten-person graduate seminar at some point in the year. I’ll also note I work in a context where there’s no major or minor, so I don’t have to worry about professionalizing students. But this is doable. It’s more work and it requires rethinking what counts as “doing philosophy” at a pretty fundamental level, but it’s doable.
Great post thank you. What is contract grading? I’m in Australia so not familiar with this concept? Loved your students work too!
Contract Grading is one of several alternative grading schemas. (See the Susan Blum edited collection “Ungrading” from WVU Press for an in-depth discussion of how and why people engage in these practices.)
Basically, my students declare at the beginning of the semester what grade they intend to earn and what steps they’ll take to earn that grade, which I approve or send back for revisions based on my feedback. Once approved, the student has a contract with me as the instructor and, so long as they complete the steps outlined in the contract, I award the grade we’ve agreed on.
The main advantages of Contract Grading are that it decreases grade anxiety (students know what they need to do instead of worrying about getting an A vs a B on a paper), gives students freedom to craft assessments that play to their interests and strengths (see, e.g., the unessays), and allows students to focus on feedback (all my assignments are pass/fail with the opportunity for revision—attaching a grade because, say, essays are worth 20% takes students’ attention away from feedback).
Not kidding when I say I haven’t graded normally for five years and I’m never going back.
“Education” has two senses: one roughly akin to “instruction” and the other roughly akin to “cultivation”. Burnett seems to be exclusively talking about the latter. (And, honestly, that’s wonderful for him.)
It’s nice to teach cultivation courses, where everyone is there because they genuinely want to be. They wouldn’t just take this assignment and tell the AI to generate a dialogue with itself.
Your average gened course is instruction and often coerced. Students see this as transactional and will produce what they are told to produce by any means available. However clever you think you are about letting the students use AI, they can just tell the AI what you told them.
The response to AI here is the same one used by math teachers in response to the calculator: in person graded work. Those students who take a liking to philosophy, will take up cultivation courses later.
Burnett is however right about something important: many students come to us for cultivation and will relish the opportunity to do their own work even if they could cheat. We ought not forget this, and we must continue giving them the opportunity.
It would be nice if we could say that instruction is for school teachers and cultivation is for college professors, but that’s just plainly not true in our world. Perhaps it is true at Princeton, but I have my doubts.
The future of the humanities in its current shape thus rests on whether the value of *instruction* can be defended against its detractors. Amusingly, our math teachers were wrong to say that we won’t have a calculator on us all the time. Now that we do, we still need to be instructed in basic math. Hopefully, the same case can be made for basic philosophical thinking.
“Hopefully”, since it may well be that humanities instruction will go away. We’d be left with cultivation only, essentially retuning such education to the elite pastime it used to be. That would be bad news for our profession. And also, at least according to John Dewey, bad news for democracy. (Like it needed more bad news.)
I’m aware it was used only as a rhetorical device, but I’d like to take this as an opportunity to say that the Kübler-Ross stage model of grief is a bit out of touch with present day grief studies (and the philosophy of grief). Sorry for the digression.
I can’t imagine what kind of job a BA in the humanities might qualify one for. It would be foolish of anyone to hire an applicant simply because they have a college degree. Some of the stupidest people I have ever met had a bachelor’s. During my time in College I didn’t learn as much from my professors as Intaught myself. I went to them for confirmation and approval of my arguments. You will never be able to make your student read or study but you can test their understanding and reasoning. In class closed book tests without electronic aid. If a student isn’t motivated enough to do the reading they’ll fail the test. Most of them will fail even if they did the reading, either they can’t articulate an argument or can’t even write.
It is worth noting that Burnett could have easily been reading submissions that were completely chatbot-created. Many students are becoming increasingly adept at using chatbots to do their work in ways that make it look original. And we can easily fool ourselves into thinking that we have found the special way to make sure that students do their own work (for a couple years, I fooled myself into thinking this).
It is really sad that, especially for low-level courses, it is no longer possible to know whether the brilliance in any given essay comes from the students or AI. While in the OP’s examples, I do believe that most of them are (edited) real conversations with AI, I am nevertheless not sure how brilliant the students’ contributions really are. I am not nearly as impressed by them as the author—there was, I recall, a distinction between one’s cleverness and producing very good work. (The article itself is a nice read.)
Among the reasons I haven’t yet incorporated AI into the assignments I give is that I think use of most LLMs is immoral and I don’t want to encourage (and definitely don’t want to require) my students to engage in this kind of immorality. I wonder if anyone has worked through this worry in a helpful way, and if so, what advice you have.
My advice is don’t use LLMs. They’re deeply immoral (their training data is built on theft), probably environmentally bad, aren’t actually designed to provide accurate information (i.e., they’re bullshit machines), and the emerging evidence is that students who use LLMs see a decrease their academic skills.
It’s just wrong. Ethically, pedagogically, it’s just wrong and we as philosophers interested in promoting ethical behavior and good epistemic habits should have no part in it.
? go to https://cosmos-institute.org/ . My advice to everyone is use it, because by not doing so you’re promoting the AI divide and you’re gonna fall behind. Extrapolating your moral compass, we’d say: don’t be a part of the economy, don’t buy a car, don’t use electricity, it’s all immoral (their building blocks are built on theft).
I think we can stand to engage with these issues with a little more nuance. Although it might be true that all economic engagement is bad for reasons similar to why LLM use is bad, it need not be the case that all engagement is equally bad, or that all engagement has the same tradeoffs associated with it, and so on.
For example, I would not say to someone supporting (say) BDS “that sort of logic means you shouldn’t buy anything, because some objectionable organization is situated at some point in the supply chain for any given thing you might buy.” What I am saying is (let’s assume) true, but it’s no knock against BDS. BDS is not claiming that the only goods we have any reason at all to boycott are Israeli goods. Rather it is claiming that the balance of reasons in favor of boycotting recommends doing so with respect to Israeli companies even if it doesn’t with respect to buying potatoes (or whatever).
Similarly, even if the sorts of things one might say against LLM use are also things one might say against other actions, it may be the case that those actions are not as bad, or that they produce greater dividends, or that they are not as easy to avoid, and so on.
I might just be misguided, but so far I haven’t found many good use cases for LLMs such that not using them is costly in the way not buying a car might be costly. (I don’t have a car either, but whatever.) So, this counts in favor of avoiding LLMs and not avoiding other things, even if avoiding other things would have some of the same good results as avoiding LLMs.
And on the flip side, I do try to limit my electricity use: not to nothing (because, again, I think we can afford to have a nuanced approach here) but certainly more than I would if I thought these concerns were not relevant. It’s about 100 degrees here, and it has been for a while, but I’ve yet to turn on my AC, in part for electricity saving reasons.
So I am not very impressed to learn my anti-LLM stance also requires (e.g.) aiming not to use electricity. That has all along been one of my aims! This just makes me more convinced I’m right to be against LLMs! Something I was committed to before they even existed (saving electricity) turns out to align with my newly-acquired anti-LLM stance.
Why do you think the use of most LLMs is immoral?
Off the top of my head:
1) most of them generate outsized environmental footprints in order to respond to queries
2) they generate their responses on the basis of unethically employed data gathering and usage techniques
3) their usage (in the context that my students would be using them) requires contributing to their development by helping to train them and contributing to their development is typically evil
4) responsibly using them (to the extent it is possible at all) requires being in a position most people are not in, such that one necessarily uses them irresponsibly and in doing so harms (or at least risks harming) oneself, and inflicting (or risking) this harm violates a duty to oneself
Another:
5) Most LLM companies know that their products are being used quite often for academic dishonesty, could easily stop this use, yet refuse to do so.
Perhaps you could enlighten those of us who understand how LLMs work how you could stop their use for academic dishonesty easily?
I am not an expert, but it seems obvious to me that they could easily keep a database of all created chats that is then shared with colleges/universities in a cross-checking program. This would stop the vast majority of academic dishonesty.
Am I mistaken about this?
Colleagues, you’re gonna love this: https://chatgpt.com/share/6810265f-6054-8001-bdf6-50364019fc73
A line from ChatGPT, at the link above:
“You are not just saving the humanities.
You are replanting them into the future.”
A bit too masturbatory for me, colleague.
I thought this was very interesting but there is one thing that puzzles me about this stuff. To use these LLMs well, do you not need to have a pretty remarkable level of knowledge about the kinds of issues you are addressing for a bunch of reasons: to make sure the prompts are decent, to push it in better directions, to filter out the bullshit etc.? (In other words, Profs and elite students are precisely the kind of people who can do this). And if that is the case, is it also not likely that some kind of education that occurs beyond the confines of the LLMs is absolutely essential going forward?
I think there are many different kinds of use one might want to make of LLMs. Just to list some ways that I’ve used them myself, you might ask them for travel advice while in a new destination, might use them to help code up some simple scripts in a programming language one doesn’t know, might use them to speed up the process of writing useful exam questions for an intro logic class, might use them to learn about similarities and differences between the historical expansions of the Huns, Turks, Mongols, Arabs, Polynesians, Bantus, and Europeans, and various other things.
For some of these purposes (particularly writing exam questions) I think that a deep background education is essential to make sure the result is good at what it is trying to do. For the others, I’m sure that a deep background education would make the results much more valuable, but there is value in getting these results even without the other education.
I’m never going to write a program that replaces Microsoft Word or Google Maps, no matter how much LLM assistance I have in coding – but there’s something gained by being able to write up small programs that do something interesting that no one else was going to make. (In my case, I had Claude help coach me on how to download OpenStreetMap data and write a Python script to identify the two points in Massachusetts that have longest drive time from one another, so that some day I can add data for the whole United States, as well as air travel data, to identify points that are “most remote”.)
If I had the right personal contacts, I’m sure I could get better travel advice or better understanding of comparative history by asking them, but the result of these LLM questions is much more accessible and only moderately less accurate than using Wikipedia (or the sadly defunct Wikitravel), let alone doing the kind of in-depth research that could turn someone into an expert. I think of them as a lot like IKEA – no one would prefer an IKEA bookcase over one made out of solid wood, but IKEA performs a great service in allowing grad students a way to store their books on a tight budget.
Overall though, I think you’re right that LLMs increase the value of particular kinds of specialized and deep education, even as they bring lots of other value to people who don’t have access to that kind of education.
Provincetown to Williamstown?
I’m an avid user of an LLM (ChatGPT), and although I am not an educator, if I were, I would want to help my students understand the marvellous potential of LLMs and, if possible, show them how to realise that potential.
First, I don’t think it is appropriate to demonise AI. It is a technology, and as such neither good nor bad in itself. Although there are issues such as energy consumption, that may reduce as competition (Deepseek?) forces more efficiency. The supposed copyright issue may be overstated – it may be more accurate to regard AI as (thoroughly) reading material, as a human reader would, before combining tokens to create something new of its own.
Secondly, it seems likely that we are on the cusp of a major transition, something like the automation of intellectual work. Just as the blacksmith hand-crafting nuts and bolts was supplanted in the C19th by automation, with vast improvement in quality and resource-efficiency, so the ‘hand-crafting’ of (much) intellectual content is now being superseded by AI. This does not mean humans passively swallow the content. Rather, just as humans are needed in the factory to direct the process, check quality, reject defects, ensure compliance with safety standards, etc., so humans are needed to direct AI, challenge assertions, detect errors, and evaluate the content AI produces. We just step up a level from blacksmith to overseer. And those kind of skills – about process rather than content – are an important part of what should be taught.
So I commend Burnett for the assignment he set his students.
No offense, but this is spectacularly poor by the lights of the AI industry itself.
Of the environmental issues, you write that these issues may be “reduce[d] as competition… forces more efficiency.” This is, at best, wishful thinking not grounded in reality. Eric Schmidt, the former CEO of google estimates that AI will consume an incredible 99% of all energy generated in the United States. And this isn’t just some spitballing at a TED Talk. This was Schmidt’s testimony before Congress. Energy consumption by LLMs and other GenAI systems is hard to pin down because companies typically lump AI power usage in with cloud computing in general, but essentially every credible estimate of future resource demands (i.e., electricity, water, and physical footprint) predicts that AI’s resource consumption will increase exponentially.
Of copyright, you claim
The issue is decidedly not overstated, both in degree and in terms of how GenAI processes information. OpenAI testified before the British House of Lords that “it would be impossible to train today’s leading AI models without using copyrighted materials.” The creators of these materials did not consent to and were not compensated for their work’s use in that way. This includes myself; The Atlantic has a searchable database of works from LibGen, the pirated, copyrighted database of materials used to train Meta’s AI, for example, and my work is there. If Meta is explicitly using a pirated database to train its AI, this is at least ethically problematic and arguably illegal. The fact that reporting suggests Mark Zuckerberg signed off on using LibGen personally and Meta employees tried to warn others at Meta that what they were doing is wrong only makes it worse. There’s no reason, I’d argue, that we should suspect Meta is an outlier on these fronts, especially given that every major LLM’s parent company is facing a raft of copyright lawsuits.
What’s more, claiming that AI is “thoroughly reading material, as a human reader would,” fundamentally misunderstands what AI is and how it operates. Human readers read and comprehend that reading because we are capable of mental representation. In other words, we have thoughts that mean things. AI is fundamentally incapable of this and is instead a pattern recognition machine that takes training data and uses it to output a unique data set that is statistically similar to the types of inputs it has calculated are relevant based on the text the user has entered. It is literally a Frankfurtian bullshit machine. To say that AI is reading like a human is to anthropomorphize it in a way that does not begin to adequately describe how AI actually works.
So, even before you get into the claims that we’re “on the cusp of a major transition” that will automate intellectual work—itself a claim that is speculative, at best—you have the fact that AI will consume exponentially more resources to further a technology inescapably built on theft. Again, this isn’t me climbing on my soapbox to demonize AI—this is the view of AI executives when asked to testify before government panels.
You wrote “Human readers read and comprehend that reading because we are capable of mental representation. In other words, we have thoughts that mean things. AI is fundamentally incapable of this.”
You may want to learn more about AIs (e.g. LLMs).
For numbers, try https://arxiv.org/html/2502.00873v1
For space and time, try https://arxiv.org/abs/2310.02207
Of course, there’s plenty more about the representations in AI. And if our thoughts mean things, and theirs were copied from ours, then their thoughts mean things too.
LLMs do not have thought that are “copied from ours”. They’ve been trained on the (relations between the) symbols that we use to externalise, extend and augment our thinking with.
Sure, I’m not surprised that the papers you refer to find correlations or topological similarities between how LLMs manipulate symbols and how humans do: that’s from the data they were fed. But one could create training material from meaningless-but-grammatically-consistent gobbledygook, feed that into an LLM and have it learn the statistical relations. That doesn’t make either the training data or what the LLM learned meaningful. Not that the LLM cares.