Back in the day, when I was a Harvard junior, Harvard Economics hired a brand-new assistant professor named Mark Watson fresh from UCSD to teach time series econometrics at the graduate level. To me back then,and, if I remember correctly, to Andrei Shleifer and Steven Kaplan as well, that seemed to be a thing worth learning. Plus by then I had clued into the fact that if I took a graduate economics course I would get an A, while if we took undergraduate economics courses there was a chance I would wind up with a B+. Incentives matter.
As I recall, back then Harvard economics graduate students had a weird culture. After the required core theory courses, they did not take, but rather audited, courses. They believed that that was the way to maximize their intellectual reach, by not spending unnecessary time doing problem sets and taking midterm exams. I think—I am not sure—that their culture held that submitting yourself to the harness of taking a course for a grade was to demonstrate that you were not up to the intellectual capabilities of your peers.
They were, of course, very wrong.
But that meant that I was among the relatively few people in the classroom doing the problem sets. Thus I was among the very few people in the classroom who understood what Mark Watson was doing at the blackboard. And so I could ask pertinent questions, and answer his. And I think he thus believed back then that I was much smarter than I am.
I learned a lot that semester, not just about the econometrics of ARIMA modeling, but meta knowledge about education and the educational process.
And here we have a very nice piece by Halina Bennet:
<https://www.slowboring.com/p/the-return-of-institutional-purpose> <http://slowboring.com>
Half a million K-8 public-school students in New York City will be banned from using generative artificial intelligence this school year, Mayor Zohran Mamdani announced on Wednesday. Meanwhile, students at Arizona Stat…
<https://www.slowboring.com/p/the-return-of-institutional-purpose> <http://slowboring.com>
- Brad DeLong here*: This is frustrating, as the entire rest of it is behind the “Slow Boring” paywall. (However, one can—I think: I do not really understand how SubStack’s Discovery Engine interacts with its Paywall Engine these days—gain access via a 7-day free trial.) But, briefly:
Before MAMLM GPT LLMs, a single written-assignment protocol (essay, problem set, exam) had bundled together many purposes at once:
transmitting content,
building discipline,
developing judgment,
performing effort,
producing a grade,
no one had to disaggregate these.
AI collapses the bundle:
a student can now generate a passable
productwithout undergoing theprocessof “learning”;severing “students who learned” from “students who appeared to learn”;
that compels institutions to figure out how to do each piece of the former bundle relevant to its purposes on its own;
or die.
MIT leads the way:
its
committee reporton A.I. use in teaching, learning, and research training is “the most exhaustive institutional self-examination on the subject… a master key that… beyond acting as a single élite institution’s…prescription… is a thorough accounting…”:it covers every aim universities might focus on,
it covers all the ways A.I.’s integration advances or undermines those aims. stated goal,
its answer is “backward design”: “start from what you want students to
know or become, then ask whether AI advances or undermines that end…”
Takeaways from Halina’s weblog post:
AI’s real intervention is epistemic, not pedagogical: destroying the
output-as-evidence-of-learningassumption;“backward design” reframes the question as “what is this education for, and does AI serve it?”;
divergent policies are entirely appropriate, aimed as they are at different pieces of the former bundle;
expect institutional sorting not just of productive cognitive struggle from job-readiness but along other axes as well
assessment goes face-to-face, handwritten-timed, and primary-source work: “the machine can assist you but cannot know your stuff for you”.
And Halina’s most important insight:
A.I. is not creating new purposes for institutions or for education on the whole. Instead, A.I. is forcing every institution at every level of education to confront the question of what its purpose actually is…
The “AI”-in-education panic is actually a governance and design problem. Institutions need to identify their values and their value propositions, and then figure out how to get students to do the work they need to do. There is no single “ban vs. embrace” war but rather different adaptations to different situations, all of which should be guided by the principal of backward design.
And so Halina runs through public K-12, vocational ed, community colleges, public multiversities, & cetera.
What do I think? Clearly, I need to read the MIT Report <https://bpb-us-e1.wpmucdn.com/sites.mit.edu/dist/d/2418/files/2026/09/AI-Committee-Final-Report-Aug-13.pdf> cover-to-cover, think about it deeply, and come back to this. That is what I think now.
That said, perhaps these are things I think. But perhaps I do not think these things—I do need to think about these issues a lot more:
The invariant hidden curriculum is the seven academic labors: survey a subject, identify live issues, hone a key question, research it, analyze evidence for an answer, store it durably, and persuade others it fits reality. Technology changes how, never what.
We’ve already, accidentally and unthinkingly, added an average internet s***poster to the course teaching staff—one with far more student contact hours than faculty..
MAMLMs are best understood as the latest abstraction layer in a five-millennia progression of the implementation of the Anthology Super-Intelligence, the real ASI, of human wisdom from from clay to papyrus to codex to press to internet.
They do two things best—serve as natural-language front-ends to un/structured data, and produce stochastic prose/code interpolations—which makes them slop machines at their worst, copilot force multipliers at their best.
They are superb survey accelerants and terrible oracles,” and that specific combination is what makes them potentially pedagogically gold and dross at once.
The genuinely big deal is that natural language interfaces enable a Sokratic, dialectical relationship with our tools that search engines and autocomplete cannot match.
Higher education’s principal purpose has been and is to create front-end nodes in the East African Plains Ape’s real ASI. That task survived Gutenberg, TV, and MOOCs. It will survive MAMLMs.
Every written assignment produces two things: the paper turned in now, and the judgment the student builds for later, and “AI” performs the first while hollowing out the second
**unless the student works not just at the “AI” abstraction layer but at the two abstraction layers beneath it as well.**There is a behavioral trap: if you work at the wrong abstraction layer, the tool’s benefit and its cost show up at different times.
The large multiversity lecture is a budgetary economizing convenience dressed as a philosophy of education.
Probably we should always have been doing the Oxbridge tutorial, which the American multiuniversity abandoned as an unscalable luxury.
The standard cannot be “write every sentence yourself”—a losing war—but “you must be able to think, in real time, about the sentences that appear under your name.”
Successful liberal-arts humanities programs will be the ones that deliver high literacy and deep numeracy—equipping students to understand and act in a largely symbolic, networked world: the crisis in the humanities is not primarily political but educational.
##public-reason
##axademia
##mamlms
#halina-bennet-the-return-of-institutional-purpose-to-academia
#halina-bennet
#the-return-of-institutional-purpose-to-academia
#institutional-purpose
#backward-design
#ai-in-education
#mit-ai-report
#oxbridge-tutorial-system
