This submit felt a bit melodramatic and/or hysterical, whilst I wrote it. However I truthfully consider each little bit of this. I’m posting it primarily in order that anybody who desires the deck and/or the essay to assist them articulate a perspective to directors and employers can have it. I don’t know if it can assist, nevertheless it’s not less than one thing to contemplate. The .tex file for the deck is on the finish additionally fwiw.
I just lately gave a chat to a small group at Baylor in regards to the wants of our college with respect to college use of AI. I talked to them below this rhetorical framing of “Learn how to Encourage Adoption of AI Amongst School”. Yow will discover the deck right here. I didn’t end my typical behavior of tweaking it till all of the Tikz was good, and kind of don’t wish to now, so it’s not completely stunning the way in which I prefer it, however the concepts are there and I needed to share them as a result of I assumed they had been value overtly speaking about and interested by. The argument I make is pretty easy, and it’s come up right here on the substack rather a lot, however right here it’s once more and roughly coalescing into one single piece.
AI Brokers are an expertise good. What do I imply? I imply ex ante an individual unfamiliar with utilizing an AI agent will be unable to cost their valuation of it. They can not accomplish that as a result of they don’t have a body of reference, and the extra extraordinarily completely different it’s, the much less doubtless that body of reference they try to make use of is even remotely correct.
What may that body of reference be? Oh I don’t know — perhaps ChatGPT? Perhaps that is simply ChatGPT proper? Or perhaps this is sort of a Nintendo? Or is it a Sony Walkman? Just like the modal tutorial not in laptop science in all probability provides two rips to be sincere. In the event that they don’t know what an AI Agent is now, in February 2026, likelihood is they perhaps weren’t even paying the $20/month for ChatGPT. Perhaps they had been utilizing the free model. Is that loopy to think about? No. It’s not loopy. It’s nearly definitely the case. Most of these weekly customers of ChatGPT will not be paying for it.
Having a PhD doesn’t imply you perceive the worth of AI, or that there’s any worth. It’s an expertise good, its worth have to be skilled first subjectively, and then you definitely’ll know over time by way of experimentation simply what it’s usefulness is and subsequently what your willingness to pay is. I made this gorgeous crappy image however you get the purpose:
AI Triggers Repugnance. It is a completely different concept, and it’s one I’ve been toying with for now 3 years. I believe synthetic intelligence, particularly the big language fashions, are for many individuals repugnant. I imply they’re morally offensive, not due to the labor substitution or the automation or the impression on the surroundings. A willingness to carry these concepts could the truth is be endogenous to some deeper nearly primal factor which is that that is know-how that talks to me and that’s so totally reprehensible that I would like nothing to do with it, nor do I would like anybody else to make use of it.
That is sort of a “repugnance as a constraint on markets” concept or the cognition of disgust. It’s not bias per se — it’s wrapped up in psychology but additionally ethics and affordable coverage making, little doubt. However I’m saying that I wager you that individuals have a special response to the truth that this software program passes the Turing take a look at so effectively. Excel might have most of the identical impacts as ChatGPT and never set off a lot ethical opposition. I believe the way in which it insists on being private and intimate because it solutions our questions and performs the duties we give it locations it firmly within the uncanny valley, however I don’t assume it actually can exit that valley, not less than not for a lot of, as it’s totally alien, inhuman, pretending to be human, doing unusual issues to us and with us that we don’t perceive, and we’re all jaded about social media and telephones anyway by now that it is sensible that there’s a lot warning, skepticism, and trepidation.
Frontier Fashions Are Costly. So, right here’s the issue. Most of our employers is not going to be paying for the very subscription tier that you just want so as to be productive utilizing Claude Code or another equal AI frontier mannequin. In contrast to ChatGPT, the place individuals might get through the use of the free model to do trivial duties, that’s unattainable with Claude Code, although there’s a free model. There’s not a activity like that the place Claude code is beneficial.
When you expertise what Claude Code can do, and also you push it to do extra, and also you problem it to do increasingly more onerous issues in your life, you can’t keep away from the unlucky reality which is that you must have it.
Safety Dangers Are Doubtless Enormous. The issue is, I don’t assume your or my college employer pays for it. They won’t pay for it as a result of the safety dangers are large.
I bear in mind as soon as instructing this class on environmental economics in grad faculty. And the writer was speaking about air pollution and smokestacks. However then he stated one thing attention-grabbing — he stated automobile drivers are like actually unhealthy, beginner smokestack operators. The corporations not less than try to be environment friendly within the surroundings they’re in for the sake of revenue maximization. However regular individuals driving their automobiles — they idle, they don’t take care of them, they simply pour poison into the air and in mixture do a whole lot of harm.
I believe most of us, reality be identified, are like automobile drivers. We’re beginner smokestack operators. We all the time assume it’s going to be another person who creates actual safety dangers for the system, however in all probability it’s us. It’s like that factor the place everyone believes they’re above common. I believe we in all probability assume we’re extra subtle and fewer susceptible to falling sufferer to malicious assaults than we in all probability are.
Which if I’m proper, signifies that the colleges are in a jam. On the one hand, lots of them are pushing (who is aware of why? Is it coming from the regents? Donors?) for college to undertake AI. However for what? And why? What’s the worth? What’s the use? What’s inappropriate use?
Effectively, the actually helpful use is excessive common reward, excessive variance. It’s like adopting a wild bull and attempting to show it to be civilized. If all you utilize AI for is writing your emails, it will possibly do no harm and it has no actual worth. However when you can harness Claude Code in direction of bettering the standard of your instructing whereas concurrently serving to you make actual discoveries, actual progress in your analysis — effectively that isn’t a free lunch. These will not be free. That’s going to price — I don’t imply price cash, however it can price cash. I imply there shall be safety dangers. There shall be vulnerabilities uncovered.
Productiveness Enhancements to Cut up the Market. So now take into account this. What if the productiveness positive factors results in a rise in provide however not a rise in demand. Larry Katz doesn’t broaden the variety of points on the QJE, nor do any of the opposite journals for that matter. The slots keep the identical. However the variety of papers develop. The variety of jobs keep the identical in academia. Actually they fall due to falling fertility that’s slamming as we communicate on the door of the 2026 getting into class and thereafter and the availability facet shocks to analysis universities all over the place from dried up federal grants and sharp reductions in overhead.
So now take into account this. Another person will get one thing that very effectively will improve the variety of papers written. Some will write worse papers than the common, like what Reimers and Waldfogel discovered with different inventive outputs (e.g., books). However this can widen the distribution, and positively improve the noise if it has no impact on one of the best papers. So what does a noisier course of that the journals face imply for you? Does it assist you to? No, it doesn’t. We’re all about to face a really sturdy wind. Perhaps the strongest one we’ve seen earlier than.
Expertise and Subsidies Are Crucial For Adoption. So that is the issue. The issue is that almost all of us can’t afford to not undertake. Most of us should not have the luxurious of not adopting. Now we have households, we’ve got desires. Now we have papers inside us that we would like out, and but we at the moment are competing with robots to write down them. Which is healthier for society? What if the robots write higher papers? What if robots write markdowns which are higher than our greatest papers?
And but the very fact is that also there may be this hole between the perceived worth of AI Brokers for analysis and instructing, the precise worth which is unknown till you utilize it lots, and the associated fee. So universities can solely get adoption to occur, exterior of self choice, by way of two levers — they have to assist college expertise the worth such that the perceived profit and the precise profit change into the identical factor. They usually should decrease the associated fee, which is unquestionably monetary and possibly could be moderated, but additionally moral/psychological and that frankly in all probability can’t.
So I’ll take them in flip. What can the colleges deal with to assist elevate the perceived profit to one thing equalling the precise worth? In brief, it have to be one thing that’s time intensive, extremely helpful, and one thing that — no offense — the common college member is horrible at.
The Making of Decks is the First Use Case. So I’ve stated this earlier than, however I simply needed to have a motive to share the deck and put all of it collectively in a single submit. I earnestly assume that if directors are wanting college to undertake AI, they have to deal with AI brokers — particularly Claude Code in my view — and get college to make their classroom decks utilizing it. If the aim is adoption, that’s what you do to make college undertake it.
The explanation I say that is easy. Decks are extraordinarily vital objects we use for instructing. Some textbooks will make the decks for us and those they make principally are horrible, however we settle for that as a result of most college will not be nice at making decks identical to they don’t seem to be nice at making exams or homework assignments. Which can be why the textbook corporations bundle these too.
Moreover, really phenomenal decks take time even in case you are good at them, which I’d say describes in all probability 1 in 100 college (and positively not this college). Most use default settings. It’s why each speak utilizing beamer appears an identical to a different.
And simply making a mediocre deck takes hours and hours of time. Simply being unhealthy at it takes so lengthy! Simply making a horrible horrible deck is so time consuming. It’s simply terrible.
And but they’re so vital. They persist, they’re handed round, they’re studied. They’re studied in all probability greater than the books themselves. They’re objects of studying. They provide confidence to the trainer. The trainer is aware of the place they’re always. It helps them talk to the category. It helps the scholars study.
When completed effectively.
What if there existed a know-how that you might use to show your lectures into one of the best decks ever made? What if they may make appropriate replacements of the decks now at no time use in any respect, and any extra time use spent on making these decks solely made the decks higher. What if there existed a know-how the place you made decks and the time you spent making the decks you had been one way or the other spending training your lecture, and studying the fabric at a deeper stage on the identical time. And what when you had a know-how that did that and but one way or the other nonetheless left you with 10 hours per week.
That and that alone is all that it’s good to get college to undertake AI. That’s it. You don’t have to workshops displaying 50 completely different use instances, new individuals coming in every week. You simply want them to make their lecture slides utilizing Claude Code and they’re going to then use it for every part else. It is going to fall like dominos.
Repair Our Analysis Directories. The opposite use case I believe I might see, however this one is a bit riskier for certain, is to have Claude Code go into our analysis and clear up the directories.
Why do I say “directories” and “analysis” as in the event that they’re the identical factor? As a result of — and right here I’m being provocative — what’s analysis precisely? Positive it’s the creation of data. It’s time spent on inventive effort and discovery. 100%. However you understand what else analysis is, much more primary than that?
Analysis is a group of folders on a pc.
And that’s the reason I believe Claude Code is definitely helpful for extra than simply utilized quantitative social scientists. Who on college will not be carrying round analysis in dropbox folders? I believe whoever that individual is doesn’t want a dropbox subscription, and might be not a candidate for experiencing the advantages of AI Brokers analysis. However when you’ve got a mission, and that mission lives in folders on a pc, then you’ll profit from Claude Code — even when I, some measly economist, can’t articulate what will probably be.
So, one of many different issues that you are able to do to assist a college member undertake AI, which would require minimizing the hole between perceived worth and precise worth through expertise, is to easily give that college this immediate:
Inform them to only kind that in. Set up Claude because the desktop app. Level it to an previous listing of theirs. In case you actually wish to freak them out — inform them to level it at their dissertation. Inform them to repeat and paste that in. Watch when one thing like this occurs.
Or have them audit code. Inform them to make use of my “referee2” code audit persona at my MixtapeTools repo to audit their previous code. Watch it because it not solely audits the code within the language you wrote it in, however replicates it in python and R too, simply to verify that no errors had been made wherever. Then writes a referee report. Then have it in a special manifestation do this report for you.
Now that is simply the beginning in fact. The true work going ahead will not be in automating all elements of analysis. The true work for us in all probability is determining how on the earth to confirm all this ourselves. Now we have to be 100% sure of every part that was completed, if it was completed. Some will really feel extra comfy with some issues greater than others, and it’s not my place on this substack to dictate that I do know the place that line is or needs to be for all individuals. I’ve an opinion, however I’m not at present able to share it. I simply am saying that whoever figures out the way to scale back errors to zero will in all probability be the best marginal product participant on this subsequent part.
Limitations, Levers and Excessive Variance Prices. However right here’s the rub. If you wish to have college undertake AI for non-trivial duties, you have to to assist them expertise what it will possibly do. And you need to scale back the prices. It’s doable experiencing it might improve AI repugnance, or it might lower it. However it can in all probability allow an individual to raised worth it, as long as they’re keen to make use of it intensively for non-trivial duties.
However then there may be the monetary prices. They don’t seem to be trivial. To do something remotely helpful with Claude Code or another AI agent will price $100 to $200/month per individual except the college is ready to enter into licensing agreements with corporations like OpenAI, Google, or Anthropic. I actually assume nobody ought to take severely the declare that there’s a fourth choice. And of these three, it’s my opinion it’s in all probability Claude Code. Which suggests there’ll should be subsidies and licenses offered, identical to there are computer systems offered to college.
And but, something that may occur, will occur, with sufficient trials. Which suggests right here that if there may be even the smallest threat of Claude Code doing one thing since you had college messing round within the terminal, and so they don’t have a clue what they’re doing, then you definitely multiply that over a thousand or extra college and much more college students, and people horrible thick tail occasions will occur — the great ones and the unhealthy ones. So this needs to be solved and those that wait to resolve it, will lose.
Conclusion. Be happy to make use of that deck. You may as well use this .tex file. You don’t want to credit score me. You don’t want to ask me for permission to make use of it. In case you assume that is useful, you should use it. I do assume these are vital and I believe that others have to discover a manner to assist get this info to directors and dept chairs.
However, I’m not personally optimistic that that is going to occur, or frankly even that it ought to occur. I believe the safety dangers are non trivial, and so I’ve purchased my very own laptop computer, and I’ve my very own subscription. I believe I simply have determined way back that I’ll put money into myself, and I don’t want others to do it on my behalf. That is my profession, that is my life, I’m the one who’s answerable for my courses, for instructing my college students, for rising as a professor, for being as inventive as I really feel I should be.
