The kids who used AI and studied for a similar amount of time as the non-AI high performers apparently had similar (slightly higher) performance. The kids who used AI and did poorly used the AI to do the homework for them (this is what the article says).
I believe AI is basically an amplifier of bad and good. I’m cynical about the world and assume it will be used more for bad than good, but I don’t doubt some of the best people in every field will be using AI to amplify their work in good ways.
It's not so clear that AI is an amplifier.. The paper has some fascinating analysis on this topic:
"At the other end of the distribution, AI students who spend more than 65 minutes on their homework receive homework and exam scores similar to those of non-AI students, suggesting that these students do not use generative AI for homework assignments. However, this group consists entirely of students who adopted generative AI no more than Öve months. Six months after adoption, no AI student spends more than 65 minutes completing their homework (see Figure A5). This is consistent with the gradual process of learning how to use AI tools. It also suggests that AI crowds out the highest level of e§ort."
"Interestingly, in the range of 50-65 minutes, the median and the interquartile range
of exam scores of AI and non-AI students are similar. This implies that, in the range
where AI students and non-AI students have overlapping homework times, students
who spend the same amount of time completing homework on average receive similar
exam scores."
"This pattern shows that students who spend the same amount of time on homework
learn similarly, with or without generative AI. In other words, generative AI reduces
time spent learning for the majority of AI students but not learning efficiency for those who spend the same time studying as the non-AI students."
Personally in my own work it's pretty obvious how it's an amplifier. You waste far less time trying to find an answer to something that confuses you.
You might say "it's good to learn research skills" and that's true to an extent, but tutors have always made people better students. And AI is a tutor you can message at any time, day or night, for free.
It would be good to see the effect of access to AI during preparation on those who previously achieved top 10% points in exams of similar topics before. I suspect they would benefit further.
My apologies for coming off as over-enthusiastic, I am currently obsessed with this study. Here is another quote:
"The negative learning effects are larger for students with higher initial achievement. The differences in the estimated full (6-10 month average) effects are substantial, with a 50% gap between the most negative effect (-24 percent) for the highest tercile and the least negative (-16 percent) for the lowest tercile. "
Not top 10% as you asked, but the closest to what you asked. My working hypothesis is that top performance is highly correlated with willingness to work hard, and AI decreases the motivation to work hard.
Interesting. Top academic performance is mostly correlated with conscientiousness (willingness to work hard and keeping track of things) and intelligence. And I'd add motivation and interest to that too.
I think if you take a physics class where the student is intelligent and intrinsically motivated through their own interest (I admit this is rare) then AI probably helps.
I think if they are intelligent, intrinsically motivated, and willing to pursue knowledge beyond what the class requires, it probably helps.
That last part is key. Intrinsic motivation doesn't mean you pursue it outside normal bounds. My kid loves soccer, its her second favorite thing in the world, she has an absurdly high tolerance for physical discomfort while playing, but she doesn't play it at home. There's other things she's rather do, such as play with her toys.
When you move the bar to something even less interesting to most kids like science, you're going to have a pretty huge falloff. You're basically selecting for kids who choose to do it in their spare time. I know a lot of smart kids (I run a boyscout troop, my wife a girlscout troop, both with lots of high achievers), and none of them do this.
Depends who's using it. Like many tools the force multiplier depends on the operator.
I'm confident it's an amplifier for people who know how learning works and already do a lot of it, successfully. However the level of "learning fluency" I'm talking about isn't reached for many until late college or grad school, and sometimes not at all. So I'm not surprised by the quoted results for 12-18 year olds.
And people who want to learn. Most teenagers lack agency in their studies. They aren't in high school because they love it but because they have no choice.
> Sometimes things are just common sense pure and simple.
The discussion is about "AI", so common sense is out the window. These people's professional reputations depend on addict-level "AI" usage remaining socially acceptable.
well most university exams are designed to measure how much you study. so we didn't really need a study to tell us, "Exams continue to measure what they are designed to measure."
they're not designed to measure general aptitude, or function as admissions criteria, or screen for job applications, or any other numerous things they are used for.
there can be many questions of pedagogy. one of them is, what do our exams measure and how do we use them? professors who say, "My exam is designed to measure who studies, not be used for all these other purposes that they are actually used for" - I don't buy it. It's the same as late night comedians saying they are not responsible for solutions, even when spending 90% of their air time making political jokes.
THIS is the pedagogical issue, that pedagogy has NEVER caught up with the scope of responsibilities. This is acute in STEM - I mean, the humanities departments are generally pretty well run, all things considered, in this regard. Generative AI is accelerating that pre-existing crisis.
The fact that exam scores are correlated with how much you study is not the same as exams only reflect how much you study. Two students who study the same amount could have very different exam scores. The reason that there still is a strong correlation between exam score and time of study is because if all other things being equal, students who studies more have higher exam scores.
i'm not saying they reflect how much you study. they reflect a lot of things, including that. but you ask the people who write the tests, they're going to say, how much you know or how much you study, but nonetheless, they are limited. i agree with you. that's part of my point.
let's imagine a different study. we instead compare AI-users and non-users on a Wechsler (IQ-adjacent) test.
overall, it would be surprising if AI usage impacted your Wechsler scores. someone has done this study and the impact is quite quite small. BUT. do we care? We don't use Wechsler scores for admissions, we don't use them for jobs, we don't use them for... are you getting it now? A Wechsler family test is measuring something real, just like a university exam measures something. But what do we USE them for? Wechsler and a typical university exam are, in some senses, EQUALLY vague in terms of their fitness for purpose for answering a question like, "should we hire this guy?"
Like there is an association between IQ and earnings but it is actually surprisingly small! There is an association with math education and earnings and it is also surprisingly small. And consider how many people get by just fine without using a single piece of math education once they have finished school - like what if maximizing your earnings isn't all that it is about? Are you getting it now?
The issue isn't the AI usage. I can find tests that are immune to AI usage. The issue is using tests for things that they are not designed for. We pick and choose, for some subtle but nonetheless pervasive cultural reasons, which tests we use for which purpose, and very frequently, not because they are calibrated for the chosen purpose. This is coming from someone who scores very well on all these tests, and have kids, so I have a very strong incentive to buy into the status quo, and I'm telling you: academic testing has been fucked up for a long, long time.
You're right that academic exams also measure something along the lines of instruction following / obedience / willingness to jump over hoops for no good reason etc. and that's often a good signal for most kinds of jobs.
I never studied in university and yet I acheived good exam scores. If you understand a topic and ave a reasonable memory and ability to apply my our understanding not much studying is required in my experience. If you don't understand the big picture then you got to laboriously keep track of and manipulate a bunch of disparate pieces.
at least in my experience in university - i didn't really ask this question, since it is obvious to me, but some students have asked it during lecture, or some instructors have volunteered the answer ahead of time - if you ask how to perform better on the exam, usually the instructors say, "here's what you should study." they never say, "know more." the thing i am talking about is consistent with the paper. really, your takeaway should be, exams can't see how much you know!
Because "know more" isn't actionable. Knowing more is achieved by studying but not necessarily more time spent studying, but well spent effort. Staring at the page for hours and saying "I don't understand" doesn't help. Solve exercise problems, explain the material to fellow students, discuss it with them, make mind maps, bullet point summaries, work through derivations step by step, etc. There are many techniques.
At the end of the day though what matters is what you know. Furthermore, if it's a serious subject, it shouldn't matter whether you learned it from this teacher or from another school and teacher, as long as your knowledge is correct. Knowing the idiosyncracies of this particular teacher should not factor into the grade. A serious subject can be learned on one continent and examined on another. Bullshit courses are all about learning pet peeves and hobby horses of a particular teacher.
I would usually say something along the lines of “everything we covered is on the table” or “everything we covered since the last exam is on the table” depending on the nature of the test. That’s the same message as “know more” but I think it sounds politer.
The scary thing is that 81% of AI users in this study were determined to be "outsourcing" their homework to the LLMs - and the rate increases the more exposure they had to LLMs.
I thought that it was generally accepted by now that homework in the volumes that it is being assigned in the modern day was not found to be beneficial in any significant way in the first place? Maybe once all students start outsourcing it to AI it might finally die like it deserves to. People these days grow up with almost no free time for themselves, it's all school, sports/extracurriculars or homework nonstop. All worker drone and no play makes for an increasingly dysfunctional society.
What I remember hearing is that exercises and practice are needed to transform words on the page to something you know and master.
What I remember being questioned is does it make sense to do those exercises as homework or would they be better in school.
Or on the flip side should school get out early like 11 or noon, like I think the german gymnasium does and have all the exercises as homework.
The US system where children get out of school at 15:30 and still have a bunch of homework seems a little lopsided someplace.
The 3 month break in-between school years is definitely questionable.
I think the answer is a complex one because it intersects with personality and neurodivergence.
Depending on who you are and your family situation any form of homework can be a real challenge. Not because of what you are studying but because of how difficult it is to sit down and do anything you aren't passionate about. Certainly anyone with an executive function disability will have that challenge.
> Depending on who you are and your family situation any form of homework can be a real challenge. Not because of what you are studying but because of how difficult it is to sit down and do anything you aren't passionate about. Certainly anyone with an executive function disability will have that challenge.
Can confirm. I probably only really did less than 25% of all homework assigned to me in my entire life.
> I thought that it was generally accepted by now that homework in the volumes that it is being assigned in the modern day was not found to be beneficial in any significant way in the first place?
Citation needed? I have no clue where you got this from. I hadn't even heard of it as a conjecture, let alone as something anyone accepted, let alone as generally accepted...
Before I dive into the paper: the claim was that some effect was generally accepted, not that a study was performed on it and that it drew a particular conclusion. Is this actually going to establish the former or are we just assuming that the existence of a study implies general acceptance of its conclusion?
How do you expect students to learn anything when:
1. They don't do any homework.
2. All the in-class time is split between the teacher babysitting and playing social worker to problem students, and lecturing, with little to no opportunity to actually practice what they've learned?
I understand that some students don't have home environments that are conductive to doing homework well. I understand that some students are enrolled in five hours a day of extracurricular university-application-padding activities. I understand that some students have incredibly poor screen discipline and impulse control.
But I don't understand that anyone has magically figured out how to teach complicated things to students, and have it stick without them spending a lot of time practicing what they are learning.
As anyone who has tried to do something hard knows, the first step to being good at something is to spend a lot of time being pretty shit at it.
A student who has written and received feedback on 500,000 written words is going to be way better at writing than that same student who wrote 50,000, just like someone who has put 5,000 hours of focused practice into playing the piano is going to be better than my dumb ass, who has only put 100 hours in.
(If you found the solution to get good at stuff without practicing it, I'd love to get good at piano without putting any homework in on it.)
All play and no work makes for an even more dysfunctional society. Kids need to do their homework, both to train their minds but also to develop integrity and work ethic. Being able to diligently work toward a distant goal is not something you're born with.
This is kind of damning, isn't it? "Similar or slightly higher" for spending the _same time_ means it's not, in fact, helping. At the same time as being extremely damaging to a significant number of students who, of course, use the AI to cheat. Because everything about it makes cheating easy.
You aren’t being cynical, you’re arguing with disingenuous entities with money on the line. We already know it’s used primarily for the negative case. Everyone who ever intended High School or College knows this.
Definitely had a sheltered public school experience, took AP stats my senior year and realized all the top students were sharing answers from an earlier period through text messages. After figuring this out I too joined in on the action, then shortly after I quickly deskilled.
Suppose it's good to learn how elastic the brain is, in both directions, at a young age where it doesn't matter.
It's a familiar story. If you copy/paste Wikipedia and turn it in, you learn nothing. If you skim Wikipedia, hit up the references, the consult other sources, then synthesize your own thought, you probably get farther faster than you would without it.
We could probably cook up thousands of examples of the same problem (YouTube DIY tutorials, GPS navigation, etc.)
The dangerous thing about amplifying is that bad people are often more willing to amplify their activities, because they don't care about the negative effects. We are seeing this now as AI companies (and large companies of all kinds) rush to secure whatever advantages they can, regardless of the negative externalities. Meanwhile people who actually care about doing the right thing get trampled.
We need to shift the incentives by adding ruinous penalties for things that are currently quite commonplace if they are done by large players. Some dude training his own AI on his own computer can scrape and train. The fine for OpenAI or Meta using a single copyrighted book without permission should be in the tens or hundreds of millions.
What we're seeing currently in our society is a "loophole inversion" where the rules have an effect mainly via their loopholes. The most profitable activity is to find loopholes and exploit them as frenetically as possible to gain as much advantage as you can before the loophole is closed, or get people hooked on the loophole so it's retroactively legalized. Entities that are big enough to do this are big enough because they have lots of money behind them. Entities doing the same kinds of things without lots of money are not really doing much harm. So the best approach is to adopt a "sliding scale" in which even tiny violations by wealthy actors result in penalties enormously greater than fairly large violations by small players.
LLMs make significant mistakes frequently and smart people have no way of judging those mistakes outside their domain expertise. They are also sycophantic and great at being an echo chamber which makes people feel smart even if they are not.
So I think the burden of proof is on you to prove that they somehow amplify intelligence, it seems highly unlikely.
Most domains have some kind of internal consistency/theory building you can do. A smart person can certainly notice inconsistencies when trying to learn something. In fact they're likely to be points of confusion that the smart person will dive into just to try to make sense of things, even if they don't suspect the LLM is at fault.
Smart people know LLMs confabulate and tell them they’re Absolutely Right! Smart people don’t want to be embarrassed by trusting the hallucination machine and revealing their gullibility to others.
Why would a smart person go to an LLM for an answer they cannot judge or test, be succeptible to flattery and sycophancy rather than picking up on the emotional manipulation and being suspicious/sceptical of the interaction, or looking for support from an echo chamber rather than a Socratic opponent?
All of those sound like flaws and defects of dumb people?
> Why would a smart person go to an LLM for an answer they cannot judge or test, be succeptible to flattery and sycophancy rather than picking up on the emotional manipulation and being suspicious/sceptical of the interaction, or looking for support from an echo chamber rather than a Socratic opponent?
Great summary of the flaws with LLM "research"/"reasoning". It's always trying to con you, and I question the literacy and intelligence of the people who can't see this.
Just one 20 GW AI data center which is being built with gas powered turbine generator power plant, will be the largest fossil fuel power plant in the world... How exactly is that the same as one coal mine
I dunno, concentrating the environmental impact of 20GW of gas based power generation would be the equivalent of 6-9 Million cars running... All in the space of a corner of a small city
It helps smart people be barely more effective and helps dumb (more importantly, people who do not value effort, people who are lazy, people who are self absorbed) people shit out endless streams of worthless tokens that can swamp out everything.
Raising the noise floor like this only makes it that much harder to find "Smart" people, which we were already doing terrible at.
I use Claude every single day, but this is such a bad tradeoff. Maybe it will help me standup a quick fix when that is needed. Maybe it can help me dig through documentation to find relevant bits and figure out the unstated assumptions underlying it. Maybe it helps me generate test cases.
Meanwhile, my day to day life is now noise. All social media is noise. All content is noise. Slop pours onto me from all directions. Writing more test cases isn't helping me.
Am I smart? Am I dumb? I don't care, right now I'm deafened
All I've noticed is AI creating a perverse incentive to make everything as complicated and bureaucratic as possible, so only the people that know how to leverage AI to cut through it ever succeed.
I'm using Claude at work myself and am impressed with the product, but notice that this is the only reason I need to use it at all. Our product pages were shit to begin with, now they're AI-generated and somehow even worse. Our procedures are incomprehensible spaghetti with enough arbitrary context switching to give a sadistic Soviet municipal administrator an erection at the thought of watching anyone try to actually follow them.
Use AI to create inefficiencies, then use AI to bypass them. Those who can't do the latter will struggle to survive.
> Raising the noise floor like this only makes it that much harder to find "Smart" people,
It does give us a new heuristic, though: people who are willing to completely cut generative AI out of their lives (cold-turkey, if you ever started using it) are a much smaller group of, predominantly thoughtful, people. You do have to give up Claude to be part of this group, but from what you say, that's no great loss, and no longer being deafened is worth it.
This has considerable advantages over conventional elitism, because the barrier-to-entry is negative in almost all cases.
The one exception I've found is assistive tech, where the state-of-the-art is so poor that vibecoded slop is genuinely an improvement over the state-of-the-art, and in many cases the tooling simply isn't available to make your own assistive tech (unless you want to bootstrap an entire networked computing environment, which isn't very helpful when you want to do your online banking and do not, in fact, work at your bank).
But there are not many principled exceptions where you could seriously argue that the trade-off is worth it. Take mathematics, for example, which we often see touted as a "good use-case" of generative AI. The primary advantage of generative AI in mathematics is being able to search though a vast corpus of ivory towers and inconsistent terminology (without proper attribution) to locate and connect ideas that can help solve problems. The deficiency this is addressing is elitism, inadequate communication, and inadequate indexing within academic mathematics. This problem is entirely created by the academic mathematicians, and has been known for nearly a century (per https://en.wikipedia.org/w/index.php?title=Nicolas_Bourbaki&...):
> Bourbaki was founded in response to the effects of the First World War which caused the death of a generation of French mathematicians; as a result, young university instructors were forced to use dated texts. While teaching at the University of Strasbourg, Henri Cartan complained to his colleague André Weil of the inadequacy of available course material, which prompted Weil to propose a meeting with others in Paris to collectively write a modern analysis textbook.
To my knowledge, this is the only organised project to clean up and improve mathematical communication. Everything else (Metamath, Mizar, AFP, Lean) is yet another ivory tower. The Wikipedia article on this topic (https://en.wikipedia.org/wiki/Mathematical_knowledge_managem...) risks deletion as non-notable, that's how little anyone's actually trying. They made their own bed, and generative AI will only provide a brief respite from having to lie in it. (I was surprised how many other "compelling" use-cases evaporated when I applied this razor to them: the sibling comment https://news.ycombinator.com/item?id=49392265 points out one such.)
Vibe-coding assistive tech which doesn't yet exist, as a temporary scaffold to improve the quality-of-life of yourself and others in a social world dominated by non-essential access barriers is, to my knowledge, the only exception to this principle that can be justified. If you treat people who make other excuses, or who don't even bother with excuses, as not worth listening to, you lose little – and doubly-so, if you make your stance clear, so that others know the "cost" of gaining your attention.
All AI did here is exposed an old problem in education. Kids are expected to get everything perfectly for the first time but a lot of them don't so the class gets easier next year to keep up pass rates.
We need to restructure the system to treat failure as a signal instead of a disaster. Grades should come from hard randomized exams with unlimited retakes so one bad day won't hurt you. Homework should be optional material for self study, evaluated by teachers if you choose to do it but never forced.
Hold back students for individual classes instead of a whole grade so failing one can't ruin your social life and teachers are more willing to do it. F students will realize they have to study, start actually learning and then pass on the second time. No big deal. It happened to my friends in college, no reason they can't do it in high schools.
Discipline is a skill and it's one you have to get from experience. If you try and force kids to study when they don't want to "for their own good" you're not actually helping them. Everyone needs to find their own path. Let people fail.
> Homework should be optional material for self study
At Caltech, homework was assigned but had no bearing on your grade. The grades were based on the midterm and final exams.
But not mastering the homework usually resulted in flunking the exams. There were "retch" sessions after each homework assignment that was staffed by a grad student, and the purpose was to help the students understand the homework problems. I knew only one person (Hal Finney) who was so smart he didn't need to do the homework.
I learned the hard way that the path to success was:
1. never miss a lecture, no matter what
2. take notes by hand during lecture
3. do the homework on time, and make sure you understand every problem. Take advantage of the retch sessions.
Likewise at Oxford - only the final exams counted (though you had to pass first year exams to continue to the second year).
But you had tutorials each week, and if your tutors thought you weren't doing enough work they could set you exams mid-course called 'penal collections' and if you failed them you could be thrown out. They were rare but definitely not unknown.
Naval Nuclear Power School had (maybe still does) this model 20 years ago as well. Pass rate for my class cohort was 28% over the entire 2 year pipeline.
We didn't call them retch sessions, and they were taught by the instructors though. We also were encouraged to peer tutor and since we were all restricted to one building the homework was always group work allowed.
Also had badges to track time spent in the building for required study hours, though some people gave up and just slept at their desks when they started sliding down the grade scale and the hours racked up.
Generally I think I did 30 hours of studying/homework (went up and down depending on what was being taught, but was around that) a week (for 12-15 hours of actual lecturing), with some of my friends putting in 50% more. Generally the only day we weren't there was Saturdays. Most of the day Sunday was usually spent in class preparing for the next week.
I'll add one more thing to this, given that my experience as a math major at SIU was that each hour of lecture was expected to carry four hours of studying: read and summarize the course's texts (additionally, for math courses, attempt the problems) before lecture.
I found that (1) I didn't need to take anywhere near as many notes during lecture and (2) I could ask way, way, way more relevant questions.
This also helped tremendously when it came to studying for the actuarial exams.
That only works with competent lecturers. I can speak for my experience at a more mid university, where that would apply to about two thirds of the classes. For the remaining third, going to the lectures felt genuinely counterproductive and you could actually feel yourself losing braincells listening to the confused nonsense or classes held in English for Erasmus students by someone who could barely speak it coherently. It was just a complete waste of already little available time.
So the endgame was figuring out what the tests in previous years looked like (cause it was likely gonna be a copy paste affair), do a targeted study run for those exercises and 9/10 you would pass.
As a student of a top-tier French Master's degree, I consulted with a teacher to deal with exhaustion and to ask for a class rescheduling for my case. Explaining my situation, the teacher looks at me and interjected:
—Waitwaitwait. You... you went to all lectures!?
And he was right. Rather than following the curriculum, I should have developed my taste for various engineering topics and only used the classes as entertainment.
I think the first semester, or at least half of that, you should go to every lecture until you develop taste for what's useful. By the time you do your master's you should have a good sense for what classes are serious and have useful lectures and where you can just cram last minute because it's a bullshit class and where you're better off studying from a textbook.
When I was there, a long time ago, exams were timed and were usually open book open note. Blue books were filled in. You were trusted to adhere by those rules, and most students did their exams in their dorm rooms.
The evidence that the students honored the rules was some exams resulted in a 50% failure rate.
As for me, I went there because I wanted to learn the material. I did not care about getting a diploma. (Mine is in the basement somewhere.) I did not take any "easy A" classes, because I wanted a return on my time and tuition investment. (Though, easy A classes were hard to find at Caltech.) I wasn't even going to attend graduation, but my parents showed up and I attended to please them.
The classes, year by year, were dependent on mastering the previous year's classes. So if you cheat with AI, you're digging yourself into a bigger and bigger hole. Caltech rewires your brain. If you don't learn the stuff, you're going to be one of those EEs who carries around a card with V=A*R, V/A=R, V/R=A printed on it.
Unlimited retakes would be an enormous amount of work for professors/TAs/teachers.
Optional homework is often a disaster. At best, students would do it right before an exam and the goal of education is not to just pass exams. They’d probably still get a lower score than if they did the homework when they were supposed to.
What I think is better is to have a due date, but just make the maximum 10% each day it is late. So after 2 days, the highest score you could receive would be 80%.
I liked that system because it gave some flexibility with deadlines while still encouraging you to turn things in on time.
What you described about exams is like how I approach most certification exams: go through a test bank of questions until I get 80% then take the real test.
The study was done in China, where making the class easier to improve pass rates isn't really a thing, as the Gaokao operates more like a stack ranking where it doesn't matter much how well you did in an absolute sense, only that not too many others who did better than you are competing for the same spot. Unlimited retakes are possible in theory, except each costs you a year of your life and except for some extreme cases of repeat test-takers, most people are going to give up after just one or two bad results. Homework is definitely not optional, but going home from school is. (Due to the hukou system, children often have to go to school far from where their parents work, so boarding schools with teacher-supervised self-study are common.) I don't see how holding back for individual classes is supposed to work considering scheduling conflicts.
> If you try and force kids to study when they don't want to "for their own good" you're not actually helping them. Everyone needs to find their own path. Let people fail.
this rhetoric is pretending to be an alternative to coercion. IMO the ideas you are talking about are well trodden and are still coercion nonetheless.
I could be mistaken but my interpretation is that "studying for your own good" is being contrasted with "studying so you don't fail". If you actually fail students, they'll have a second motivation beyond trusting some authority figure's advice. Namely consequences.
You are missing the point. Getting a certificate for all of the programs you passed and still graduating could completely change the course of a person’s life. Community college has certificates and associates degrees to let people start working when they reach a practical limit in their general education.
It's all cute but naively glances over what this all boils to - competition.
Education system, contrary to popular belief/name, isn't tailored to educate but to select winners and losers which then will be picked on the job market.
That's why it seems absurd when you think about it as an institution that aims to educate. That's because that isn't the real purpose of it. The purpose is to stratify and classify early.
Schools do not operate in a vacuum; they serve as credentialing gatekeepers for a hyper-competitive capitalist job market. If everyone could easily retake exams until they got an A, grades would lose their primary utility for employers and universities: differentiation. Society relies on schools to provide a neat hierarchy of candidates that for one reason or another thrived in difficult environment of adolescent schooling.
The system often prioritizes compliance, endurance of boredom, and social maneuvering over actual critical thinking precisely because those traits align with corporate hierarchies.
Moneyshot: "The results are eye-opening. After six months, pupils using ai saw their average homework score rise by 18% across all subjects. The time they took to complete each assignment fell from an average of 64 minutes to 45. But come exam time, the same students scored 20% below their classmates who had not called on ai’s help. Homework scores once predicted exam performance; now those who score highest are, perversely, more likely to do worse in exams."
Study design: "David Stromberg of Stockholm University and Victor Lei and Wu Yanhui of the University of Hong Kong set out to fill the gap. They tracked 27,000 pupils aged 12-18 in China, where ai adoption has been fast. Around 80% reported using models such as Doubao and DeepSeek; the other 20% formed the control group."
As with most training the journey is the point, not the destination.
That said, I think smart use of AI could help. It could explain concepts in a way that might help you understand better, it could probe your knowledge in a more dynamic way by tailoring questions, and so on. This requires the AI be restrained by some harness, not free to write down the answers for you.
When I was in college, we had office hours with the prof and the TAs, we had group study sessions, and private tutors. Or you could ask your friends. All of those required going somewhere at a specific place and time and asking someone else to give up their time for you.
AI gives you a way to get the same help without asking another human. Say of that what you will, but not everyone was comfortable asking other humans for help even back then.
Back when I was studying there were topics which I had a really hard time grasping. Having a personal tutor that I could ask specific questions and have a back and forth with would have helped a lot I think. At least the few times I did have such an opportunity that was definitely the case.
Kids are, by and large, lazy however. Hell, adults are lazy, this isn't even an indictment on kids or even people in general, I think our bodies & brains are simply wired to seek the path of least resistance from an evolutionary POV.
Kids won't be using LLM tooling as a personal tutor following some sort of Socratic method, they'll ask it to solve their home/coursework for them and blindly copy/paste the answer. Hell, they'll just manually copy down what's on their screen if copy/pasting isn't possible for whatever reason.
Obviously exceptions exist, but I'd wager from being an ex-kid myself the type of kid who would genuinely use these tools for actual proper self-tutoring would be an extreme rarity.
I'm way beyond homework, but I do appreciate the compliment on seeming young!
I meant: what's going to stop the average lazy student from ignoring whatever harness the university recommends and instead use an unrestricted LLM, thus learning nothing?
> what's going to stop the average lazy student from ignoring whatever harness the university recommends and instead use an unrestricted LLM, thus learning nothing?
Nothing until they start failing exams and maybe seeing the error of their ways. My suggestions wasn't for you, it was for the smart student who wants to use AI to enhance their learning but not have it do all the work.
But cheating isn't new. People have been cheating in school for centuries. It's easier now, but the consequences were always the same. At some point the chickens come home to roost and you pay the piper.
> But cheating isn't new. People have been cheating in school for centuries. It's easier now, but the consequences were always the same. At some point the chickens come home to roost and you pay the piper.
My concern is that with LLMs and above average bullshitting abilities, the chickens might never come home to roost.
Using the forklift as a spotter and to assist in loading weights increased gains. Then again, a human can do all those things, and provide real human connection.
I am waiting for AI Viva Voce (“AVV”). I am really surprised we aren’t hearing more about this idea?
Try it for yourself by making a prompt like this (adapt as needed):
Pretend I am an undergraduate student of Computer Science. I am learning about early microprocessors from the 1970s. I want you to ask me an examination question as if you were doing a viva voce exam with me, to test my understanding of concepts. I want you to receive my answer and then based on what I said I want you to ask me a more specific question to probe my understanding. Repeat this interaction up to 5 times. Then grade my understanding so far, by giving me a pass, merit, credit, or distinction. Can you explain how you arrive at the grade based on my answers and your expectation of undergraduate knowledge of microprocessor theory?
I've tried stuff like this, but the problem is that you don't know if it's answers are correct, and more importantly, you don't know if its questions are on base or not. Plus its too easy to go off on tangents, especially when you ask it questions to clarify.
You can get some more success by writing down a framework beforehand, but openended dialog is still not a good way to learn with AI, from my experience.
I mean, everyone I know has been using socratic dialogue for months to aid understanding, this seems like a variation of that. The issue is that if the answer is too easy to obtain it's hard to have the discipline not to cheat.
If you structure learning in such a way that makes learning just a means to some end, and overindex on that end being the ultimate goal, that's what you get.
This is a pedagogical problem that AI merely exposed. Educators need to figure out How to make students choose the scenic route instead of having them optimize for the most efficient completion of a task.
School has always been about creating compliant workers, not educating people. They will double down on "performing the right things" and "morals" while the economy will continue on its K shaped path as AI gets more capable.
Yea, also the programs are too different. My psychology bachelor felt like a tea party. My computer science master was 60 hours per week during the hardest courses (which were supposed to go for 20 hours per week - so I could only do one course if it was at this level). My game design master felt more like we were dumped into an art school masquerading as a psychology/CS master but it really was art school. My information science bachelor felt like the only "normal" study program.
All of this was at university of which 3 of them were at the same university.
Especially the artsy game design program definitely did not feel like it was preparing me to be a cog in some giant corporate wheel.
If others studies tend toward this result, a question comes to my mind: it is a problem with AI or our educational system? or both? I would like to have some comments...
Glad to have scientific results on this, though in my view you could get this from first principles. Homework is onerous but it forces you to learn the material and get it into a configuration that works in your head, which you then validate with the exam.
as a fun aside here from the abstract it does come down to how you use it:
> AI users who maintain similar homework completion time as non-AI users experience small learning losses.
If the pure completion time of homework goes drops significantly with an LLM tool, I suspect these students are spending the extra time turning the material over in different ways to internalize, which is interesting.
I know this study is focused in China but one thing I'd like to understand is the nuance in what kind of student is likely to always reach for ai for homework help as I think (especially in the US, can't speak for other countries) the varying powers that be that can determine the quality and type of education you can receive (does your country have means to buy textbooks per student or shared, do you have funds to give students ipads to take home or no, what sets the bar on how far curriculums can go, are teacher and faculty pay and incentives simply tied to student pass rates, etc.) and those do more in shaping what "use AI responsibly" will look like and why it's so different.
Honestly I wouldn't demonize AI in education. When we were students we were forbidden from using calculators, then the internet and now it's become the norm. It's the same with AI, I think. AI has already become a part of our lives. The important thing is just teaching children how to use it correctly.
I used to skip homework in high school because it was boring and cut into my personal life outside of school (including my job) but I would always do very well on tests.
Then I stopped caring about tests too.
I did the same thing through college. The only reason they passed me and I got a degree is that I built the school’s website and I built personal ecommerce sites for the head of art and his wife to sell their paintings.
I haven’t done shit since like 7th grade.
Worked at Facebook, Apple, Microsoft, same behavior there - did basically nothing for them while making thousands off my games on App Store.
When you get out into the real world, what will matter is how effectively you use AI. When you are done with school, you can take the "exam" using AI. The problem is that teaching is trying to teach skills that are no longer relevant and is always behind what the latest going on in the industry. Every time without fail you got interns, who after two years at stanford would learn more in the 3 months on a google internship than in those two years at stanford. They learn useless shit.
My own experience is that college is fun. You can spend the time wisely, but school stuff, is not what makes you money.
Make student loans dischargeable in bankruptcy and make colleges under write them. The problem will get solved quickly.
Way before AI, the problem was very similar - universities do not teach practically useful stuff. Its been a permanent complaint about the education system, really, at least for the last 50 years.
AI just gave the students a way to dodge the slog.
But university was never intended to teach the bleeding edge. That would really be impossible in practice. Pre-phd, it is supposed to teach ways to efficiently attack a problem. you can take a bunch of "play courses", and succeeding at any of those requires pretty much only that one skill.
Nowadays, the challenge they face, is to keep teaching "problem attack methods" in a way that can't be trivialized by AI.
Though fundamentally, if you go to university, and evade learning the one thing you can learn there - thats your loss.
The kids who used AI and studied for a similar amount of time as the non-AI high performers apparently had similar (slightly higher) performance. The kids who used AI and did poorly used the AI to do the homework for them (this is what the article says).
I believe AI is basically an amplifier of bad and good. I’m cynical about the world and assume it will be used more for bad than good, but I don’t doubt some of the best people in every field will be using AI to amplify their work in good ways.
It's not so clear that AI is an amplifier.. The paper has some fascinating analysis on this topic:
"At the other end of the distribution, AI students who spend more than 65 minutes on their homework receive homework and exam scores similar to those of non-AI students, suggesting that these students do not use generative AI for homework assignments. However, this group consists entirely of students who adopted generative AI no more than Öve months. Six months after adoption, no AI student spends more than 65 minutes completing their homework (see Figure A5). This is consistent with the gradual process of learning how to use AI tools. It also suggests that AI crowds out the highest level of e§ort."
"Interestingly, in the range of 50-65 minutes, the median and the interquartile range of exam scores of AI and non-AI students are similar. This implies that, in the range where AI students and non-AI students have overlapping homework times, students who spend the same amount of time completing homework on average receive similar exam scores."
"This pattern shows that students who spend the same amount of time on homework learn similarly, with or without generative AI. In other words, generative AI reduces time spent learning for the majority of AI students but not learning efficiency for those who spend the same time studying as the non-AI students."
Personally in my own work it's pretty obvious how it's an amplifier. You waste far less time trying to find an answer to something that confuses you.
You might say "it's good to learn research skills" and that's true to an extent, but tutors have always made people better students. And AI is a tutor you can message at any time, day or night, for free.
Right I was not clear enough about separating my opinion from the findings.
It would be good to see the effect of access to AI during preparation on those who previously achieved top 10% points in exams of similar topics before. I suspect they would benefit further.
My apologies for coming off as over-enthusiastic, I am currently obsessed with this study. Here is another quote:
"The negative learning effects are larger for students with higher initial achievement. The differences in the estimated full (6-10 month average) effects are substantial, with a 50% gap between the most negative effect (-24 percent) for the highest tercile and the least negative (-16 percent) for the lowest tercile. "
Not top 10% as you asked, but the closest to what you asked. My working hypothesis is that top performance is highly correlated with willingness to work hard, and AI decreases the motivation to work hard.
Interesting. Top academic performance is mostly correlated with conscientiousness (willingness to work hard and keeping track of things) and intelligence. And I'd add motivation and interest to that too.
I think if you take a physics class where the student is intelligent and intrinsically motivated through their own interest (I admit this is rare) then AI probably helps.
I think if they are intelligent, intrinsically motivated, and willing to pursue knowledge beyond what the class requires, it probably helps.
That last part is key. Intrinsic motivation doesn't mean you pursue it outside normal bounds. My kid loves soccer, its her second favorite thing in the world, she has an absurdly high tolerance for physical discomfort while playing, but she doesn't play it at home. There's other things she's rather do, such as play with her toys.
When you move the bar to something even less interesting to most kids like science, you're going to have a pretty huge falloff. You're basically selecting for kids who choose to do it in their spare time. I know a lot of smart kids (I run a boyscout troop, my wife a girlscout troop, both with lots of high achievers), and none of them do this.
Depends who's using it. Like many tools the force multiplier depends on the operator.
I'm confident it's an amplifier for people who know how learning works and already do a lot of it, successfully. However the level of "learning fluency" I'm talking about isn't reached for many until late college or grad school, and sometimes not at all. So I'm not surprised by the quoted results for 12-18 year olds.
> Depends who's using it. Like many tools the force multiplier depends on the operator.
How would you prove/disprove this assumption without falling into a True Scotsman fallacy?
And people who want to learn. Most teenagers lack agency in their studies. They aren't in high school because they love it but because they have no choice.
Sometimes things are just common sense pure and simple.
> Sometimes things are just common sense pure and simple.
The discussion is about "AI", so common sense is out the window. These people's professional reputations depend on addict-level "AI" usage remaining socially acceptable.
well most university exams are designed to measure how much you study. so we didn't really need a study to tell us, "Exams continue to measure what they are designed to measure."
they're not designed to measure general aptitude, or function as admissions criteria, or screen for job applications, or any other numerous things they are used for.
there can be many questions of pedagogy. one of them is, what do our exams measure and how do we use them? professors who say, "My exam is designed to measure who studies, not be used for all these other purposes that they are actually used for" - I don't buy it. It's the same as late night comedians saying they are not responsible for solutions, even when spending 90% of their air time making political jokes.
THIS is the pedagogical issue, that pedagogy has NEVER caught up with the scope of responsibilities. This is acute in STEM - I mean, the humanities departments are generally pretty well run, all things considered, in this regard. Generative AI is accelerating that pre-existing crisis.
The fact that exam scores are correlated with how much you study is not the same as exams only reflect how much you study. Two students who study the same amount could have very different exam scores. The reason that there still is a strong correlation between exam score and time of study is because if all other things being equal, students who studies more have higher exam scores.
i'm not saying they reflect how much you study. they reflect a lot of things, including that. but you ask the people who write the tests, they're going to say, how much you know or how much you study, but nonetheless, they are limited. i agree with you. that's part of my point.
let's imagine a different study. we instead compare AI-users and non-users on a Wechsler (IQ-adjacent) test.
overall, it would be surprising if AI usage impacted your Wechsler scores. someone has done this study and the impact is quite quite small. BUT. do we care? We don't use Wechsler scores for admissions, we don't use them for jobs, we don't use them for... are you getting it now? A Wechsler family test is measuring something real, just like a university exam measures something. But what do we USE them for? Wechsler and a typical university exam are, in some senses, EQUALLY vague in terms of their fitness for purpose for answering a question like, "should we hire this guy?"
Like there is an association between IQ and earnings but it is actually surprisingly small! There is an association with math education and earnings and it is also surprisingly small. And consider how many people get by just fine without using a single piece of math education once they have finished school - like what if maximizing your earnings isn't all that it is about? Are you getting it now?
The issue isn't the AI usage. I can find tests that are immune to AI usage. The issue is using tests for things that they are not designed for. We pick and choose, for some subtle but nonetheless pervasive cultural reasons, which tests we use for which purpose, and very frequently, not because they are calibrated for the chosen purpose. This is coming from someone who scores very well on all these tests, and have kids, so I have a very strong incentive to buy into the status quo, and I'm telling you: academic testing has been fucked up for a long, long time.
You're right that academic exams also measure something along the lines of instruction following / obedience / willingness to jump over hoops for no good reason etc. and that's often a good signal for most kinds of jobs.
I never studied in university and yet I acheived good exam scores. If you understand a topic and ave a reasonable memory and ability to apply my our understanding not much studying is required in my experience. If you don't understand the big picture then you got to laboriously keep track of and manipulate a bunch of disparate pieces.
> well most university exams are designed to measure how much you study.
Huh? They're designed to measure how much you know. They can't see how much you study, nor would they have reason to be interested.
How much you know at the moment of the exam. Not how much you truly know.
at least in my experience in university - i didn't really ask this question, since it is obvious to me, but some students have asked it during lecture, or some instructors have volunteered the answer ahead of time - if you ask how to perform better on the exam, usually the instructors say, "here's what you should study." they never say, "know more." the thing i am talking about is consistent with the paper. really, your takeaway should be, exams can't see how much you know!
Because "know more" isn't actionable. Knowing more is achieved by studying but not necessarily more time spent studying, but well spent effort. Staring at the page for hours and saying "I don't understand" doesn't help. Solve exercise problems, explain the material to fellow students, discuss it with them, make mind maps, bullet point summaries, work through derivations step by step, etc. There are many techniques.
At the end of the day though what matters is what you know. Furthermore, if it's a serious subject, it shouldn't matter whether you learned it from this teacher or from another school and teacher, as long as your knowledge is correct. Knowing the idiosyncracies of this particular teacher should not factor into the grade. A serious subject can be learned on one continent and examined on another. Bullshit courses are all about learning pet peeves and hobby horses of a particular teacher.
You become a person who knows more, by studying more about the things you need to know.
I would usually say something along the lines of “everything we covered is on the table” or “everything we covered since the last exam is on the table” depending on the nature of the test. That’s the same message as “know more” but I think it sounds politer.
The scary thing is that 81% of AI users in this study were determined to be "outsourcing" their homework to the LLMs - and the rate increases the more exposure they had to LLMs.
The "slightly higher" performance is based on statistically insignificant samples (between 4 and 20 students, depending on the context, out of the total population of 26,000): https://bsky.app/profile/benjaminjriley.bsky.social/post/3mt...
I thought that it was generally accepted by now that homework in the volumes that it is being assigned in the modern day was not found to be beneficial in any significant way in the first place? Maybe once all students start outsourcing it to AI it might finally die like it deserves to. People these days grow up with almost no free time for themselves, it's all school, sports/extracurriculars or homework nonstop. All worker drone and no play makes for an increasingly dysfunctional society.
What I remember hearing is that exercises and practice are needed to transform words on the page to something you know and master.
What I remember being questioned is does it make sense to do those exercises as homework or would they be better in school.
Or on the flip side should school get out early like 11 or noon, like I think the german gymnasium does and have all the exercises as homework.
The US system where children get out of school at 15:30 and still have a bunch of homework seems a little lopsided someplace.
The 3 month break in-between school years is definitely questionable.
I think the answer is a complex one because it intersects with personality and neurodivergence.
Depending on who you are and your family situation any form of homework can be a real challenge. Not because of what you are studying but because of how difficult it is to sit down and do anything you aren't passionate about. Certainly anyone with an executive function disability will have that challenge.
> Depending on who you are and your family situation any form of homework can be a real challenge. Not because of what you are studying but because of how difficult it is to sit down and do anything you aren't passionate about. Certainly anyone with an executive function disability will have that challenge.
Can confirm. I probably only really did less than 25% of all homework assigned to me in my entire life.
> I thought that it was generally accepted by now that homework in the volumes that it is being assigned in the modern day was not found to be beneficial in any significant way in the first place?
Citation needed? I have no clue where you got this from. I hadn't even heard of it as a conjecture, let alone as something anyone accepted, let alone as generally accepted...
https://www.tandfonline.com/doi/pdf/10.1080/00220973.2012.74...
Before I dive into the paper: the claim was that some effect was generally accepted, not that a study was performed on it and that it drew a particular conclusion. Is this actually going to establish the former or are we just assuming that the existence of a study implies general acceptance of its conclusion?
How do you expect students to learn anything when:
1. They don't do any homework.
2. All the in-class time is split between the teacher babysitting and playing social worker to problem students, and lecturing, with little to no opportunity to actually practice what they've learned?
I understand that some students don't have home environments that are conductive to doing homework well. I understand that some students are enrolled in five hours a day of extracurricular university-application-padding activities. I understand that some students have incredibly poor screen discipline and impulse control.
But I don't understand that anyone has magically figured out how to teach complicated things to students, and have it stick without them spending a lot of time practicing what they are learning.
As anyone who has tried to do something hard knows, the first step to being good at something is to spend a lot of time being pretty shit at it.
A student who has written and received feedback on 500,000 written words is going to be way better at writing than that same student who wrote 50,000, just like someone who has put 5,000 hours of focused practice into playing the piano is going to be better than my dumb ass, who has only put 100 hours in.
(If you found the solution to get good at stuff without practicing it, I'd love to get good at piano without putting any homework in on it.)
> I'd love to get good at piano without putting any homework in on it
I'm pretty sure there is a way to fine-tune this [0] to auto-complete away any of your hesitations or mistakes :)
[0] https://news.ycombinator.com/item?id=49373456
All play and no work makes for an even more dysfunctional society. Kids need to do their homework, both to train their minds but also to develop integrity and work ethic. Being able to diligently work toward a distant goal is not something you're born with.
This is kind of damning, isn't it? "Similar or slightly higher" for spending the _same time_ means it's not, in fact, helping. At the same time as being extremely damaging to a significant number of students who, of course, use the AI to cheat. Because everything about it makes cheating easy.
You aren’t being cynical, you’re arguing with disingenuous entities with money on the line. We already know it’s used primarily for the negative case. Everyone who ever intended High School or College knows this.
Definitely had a sheltered public school experience, took AP stats my senior year and realized all the top students were sharing answers from an earlier period through text messages. After figuring this out I too joined in on the action, then shortly after I quickly deskilled.
Suppose it's good to learn how elastic the brain is, in both directions, at a young age where it doesn't matter.
> at a young age where it doesn't matter.
Doesn't it matter the most at a young age?
> I believe AI is basically an amplifier of bad and good.
Same can be said of technology in general tbh.
It's a familiar story. If you copy/paste Wikipedia and turn it in, you learn nothing. If you skim Wikipedia, hit up the references, the consult other sources, then synthesize your own thought, you probably get farther faster than you would without it.
We could probably cook up thousands of examples of the same problem (YouTube DIY tutorials, GPS navigation, etc.)
The dangerous thing about amplifying is that bad people are often more willing to amplify their activities, because they don't care about the negative effects. We are seeing this now as AI companies (and large companies of all kinds) rush to secure whatever advantages they can, regardless of the negative externalities. Meanwhile people who actually care about doing the right thing get trampled.
We need to shift the incentives by adding ruinous penalties for things that are currently quite commonplace if they are done by large players. Some dude training his own AI on his own computer can scrape and train. The fine for OpenAI or Meta using a single copyrighted book without permission should be in the tens or hundreds of millions.
What we're seeing currently in our society is a "loophole inversion" where the rules have an effect mainly via their loopholes. The most profitable activity is to find loopholes and exploit them as frenetically as possible to gain as much advantage as you can before the loophole is closed, or get people hooked on the loophole so it's retroactively legalized. Entities that are big enough to do this are big enough because they have lots of money behind them. Entities doing the same kinds of things without lots of money are not really doing much harm. So the best approach is to adopt a "sliding scale" in which even tiny violations by wealthy actors result in penalties enormously greater than fairly large violations by small players.
This is my thoughts as well! It makes smarter people smarter and dumb people dumber!
There is zero evidence for this.
LLMs make significant mistakes frequently and smart people have no way of judging those mistakes outside their domain expertise. They are also sycophantic and great at being an echo chamber which makes people feel smart even if they are not.
So I think the burden of proof is on you to prove that they somehow amplify intelligence, it seems highly unlikely.
Most domains have some kind of internal consistency/theory building you can do. A smart person can certainly notice inconsistencies when trying to learn something. In fact they're likely to be points of confusion that the smart person will dive into just to try to make sense of things, even if they don't suspect the LLM is at fault.
No evidence but somewhat of a counterpoints:
Smart people know LLMs confabulate and tell them they’re Absolutely Right! Smart people don’t want to be embarrassed by trusting the hallucination machine and revealing their gullibility to others.
Why would a smart person go to an LLM for an answer they cannot judge or test, be succeptible to flattery and sycophancy rather than picking up on the emotional manipulation and being suspicious/sceptical of the interaction, or looking for support from an echo chamber rather than a Socratic opponent?
All of those sound like flaws and defects of dumb people?
> Why would a smart person go to an LLM for an answer they cannot judge or test, be succeptible to flattery and sycophancy rather than picking up on the emotional manipulation and being suspicious/sceptical of the interaction, or looking for support from an echo chamber rather than a Socratic opponent?
Great summary of the flaws with LLM "research"/"reasoning". It's always trying to con you, and I question the literacy and intelligence of the people who can't see this.
All at the cost of ... {List of negatives regarding the construction and powering of AI data centers }
If you added up the negative effect of every single "ai data center" you'd end up with less than a single coal mine.
It's good to think about second order impacts but this is not useful.
Citation needed,
Just one 20 GW AI data center which is being built with gas powered turbine generator power plant, will be the largest fossil fuel power plant in the world... How exactly is that the same as one coal mine
The social costs are worse.
I dunno, concentrating the environmental impact of 20GW of gas based power generation would be the equivalent of 6-9 Million cars running... All in the space of a corner of a small city
That sounds much worse than I had expected. Is this enough consumption to significantly affect the local price of gasoline?
Edit: Why don't they have lower-emission power sources?
Most importantly, it makes investors think dumb people are smart.
It helps smart people be barely more effective and helps dumb (more importantly, people who do not value effort, people who are lazy, people who are self absorbed) people shit out endless streams of worthless tokens that can swamp out everything.
Raising the noise floor like this only makes it that much harder to find "Smart" people, which we were already doing terrible at.
I use Claude every single day, but this is such a bad tradeoff. Maybe it will help me standup a quick fix when that is needed. Maybe it can help me dig through documentation to find relevant bits and figure out the unstated assumptions underlying it. Maybe it helps me generate test cases.
Meanwhile, my day to day life is now noise. All social media is noise. All content is noise. Slop pours onto me from all directions. Writing more test cases isn't helping me.
Am I smart? Am I dumb? I don't care, right now I'm deafened
All I've noticed is AI creating a perverse incentive to make everything as complicated and bureaucratic as possible, so only the people that know how to leverage AI to cut through it ever succeed.
I'm using Claude at work myself and am impressed with the product, but notice that this is the only reason I need to use it at all. Our product pages were shit to begin with, now they're AI-generated and somehow even worse. Our procedures are incomprehensible spaghetti with enough arbitrary context switching to give a sadistic Soviet municipal administrator an erection at the thought of watching anyone try to actually follow them.
Use AI to create inefficiencies, then use AI to bypass them. Those who can't do the latter will struggle to survive.
> Raising the noise floor like this only makes it that much harder to find "Smart" people, which we were already doing terrible at.
Clarification: to value “smart” people, which we were already doing terrible at.
> Raising the noise floor like this only makes it that much harder to find "Smart" people,
It does give us a new heuristic, though: people who are willing to completely cut generative AI out of their lives (cold-turkey, if you ever started using it) are a much smaller group of, predominantly thoughtful, people. You do have to give up Claude to be part of this group, but from what you say, that's no great loss, and no longer being deafened is worth it.
This has considerable advantages over conventional elitism, because the barrier-to-entry is negative in almost all cases.
The one exception I've found is assistive tech, where the state-of-the-art is so poor that vibecoded slop is genuinely an improvement over the state-of-the-art, and in many cases the tooling simply isn't available to make your own assistive tech (unless you want to bootstrap an entire networked computing environment, which isn't very helpful when you want to do your online banking and do not, in fact, work at your bank).
But there are not many principled exceptions where you could seriously argue that the trade-off is worth it. Take mathematics, for example, which we often see touted as a "good use-case" of generative AI. The primary advantage of generative AI in mathematics is being able to search though a vast corpus of ivory towers and inconsistent terminology (without proper attribution) to locate and connect ideas that can help solve problems. The deficiency this is addressing is elitism, inadequate communication, and inadequate indexing within academic mathematics. This problem is entirely created by the academic mathematicians, and has been known for nearly a century (per https://en.wikipedia.org/w/index.php?title=Nicolas_Bourbaki&...):
> Bourbaki was founded in response to the effects of the First World War which caused the death of a generation of French mathematicians; as a result, young university instructors were forced to use dated texts. While teaching at the University of Strasbourg, Henri Cartan complained to his colleague André Weil of the inadequacy of available course material, which prompted Weil to propose a meeting with others in Paris to collectively write a modern analysis textbook.
To my knowledge, this is the only organised project to clean up and improve mathematical communication. Everything else (Metamath, Mizar, AFP, Lean) is yet another ivory tower. The Wikipedia article on this topic (https://en.wikipedia.org/wiki/Mathematical_knowledge_managem...) risks deletion as non-notable, that's how little anyone's actually trying. They made their own bed, and generative AI will only provide a brief respite from having to lie in it. (I was surprised how many other "compelling" use-cases evaporated when I applied this razor to them: the sibling comment https://news.ycombinator.com/item?id=49392265 points out one such.)
Vibe-coding assistive tech which doesn't yet exist, as a temporary scaffold to improve the quality-of-life of yourself and others in a social world dominated by non-essential access barriers is, to my knowledge, the only exception to this principle that can be justified. If you treat people who make other excuses, or who don't even bother with excuses, as not worth listening to, you lose little – and doubly-so, if you make your stance clear, so that others know the "cost" of gaining your attention.
All AI did here is exposed an old problem in education. Kids are expected to get everything perfectly for the first time but a lot of them don't so the class gets easier next year to keep up pass rates.
We need to restructure the system to treat failure as a signal instead of a disaster. Grades should come from hard randomized exams with unlimited retakes so one bad day won't hurt you. Homework should be optional material for self study, evaluated by teachers if you choose to do it but never forced.
Hold back students for individual classes instead of a whole grade so failing one can't ruin your social life and teachers are more willing to do it. F students will realize they have to study, start actually learning and then pass on the second time. No big deal. It happened to my friends in college, no reason they can't do it in high schools.
Discipline is a skill and it's one you have to get from experience. If you try and force kids to study when they don't want to "for their own good" you're not actually helping them. Everyone needs to find their own path. Let people fail.
> Homework should be optional material for self study
At Caltech, homework was assigned but had no bearing on your grade. The grades were based on the midterm and final exams.
But not mastering the homework usually resulted in flunking the exams. There were "retch" sessions after each homework assignment that was staffed by a grad student, and the purpose was to help the students understand the homework problems. I knew only one person (Hal Finney) who was so smart he didn't need to do the homework.
I learned the hard way that the path to success was:
1. never miss a lecture, no matter what
2. take notes by hand during lecture
3. do the homework on time, and make sure you understand every problem. Take advantage of the retch sessions.
And that worked for me.
Likewise at Oxford - only the final exams counted (though you had to pass first year exams to continue to the second year).
But you had tutorials each week, and if your tutors thought you weren't doing enough work they could set you exams mid-course called 'penal collections' and if you failed them you could be thrown out. They were rare but definitely not unknown.
Naval Nuclear Power School had (maybe still does) this model 20 years ago as well. Pass rate for my class cohort was 28% over the entire 2 year pipeline.
We didn't call them retch sessions, and they were taught by the instructors though. We also were encouraged to peer tutor and since we were all restricted to one building the homework was always group work allowed.
Also had badges to track time spent in the building for required study hours, though some people gave up and just slept at their desks when they started sliding down the grade scale and the hours racked up.
Generally I think I did 30 hours of studying/homework (went up and down depending on what was being taught, but was around that) a week (for 12-15 hours of actual lecturing), with some of my friends putting in 50% more. Generally the only day we weren't there was Saturdays. Most of the day Sunday was usually spent in class preparing for the next week.
I've given this advice many times, and the ones who followed it did well. I only added "leave your laptop in your dorm room" for modern times.
The general pattern at Caltech was 2 hours of study for every hour of lecture. Which was quite a shock to me.
I'll add one more thing to this, given that my experience as a math major at SIU was that each hour of lecture was expected to carry four hours of studying: read and summarize the course's texts (additionally, for math courses, attempt the problems) before lecture.
I found that (1) I didn't need to take anywhere near as many notes during lecture and (2) I could ask way, way, way more relevant questions.
This also helped tremendously when it came to studying for the actuarial exams.
That only works with competent lecturers. I can speak for my experience at a more mid university, where that would apply to about two thirds of the classes. For the remaining third, going to the lectures felt genuinely counterproductive and you could actually feel yourself losing braincells listening to the confused nonsense or classes held in English for Erasmus students by someone who could barely speak it coherently. It was just a complete waste of already little available time.
So the endgame was figuring out what the tests in previous years looked like (cause it was likely gonna be a copy paste affair), do a targeted study run for those exercises and 9/10 you would pass.
> 1. never miss a lecture, no matter what
As a student of a top-tier French Master's degree, I consulted with a teacher to deal with exhaustion and to ask for a class rescheduling for my case. Explaining my situation, the teacher looks at me and interjected:
—Waitwaitwait. You... you went to all lectures!?
And he was right. Rather than following the curriculum, I should have developed my taste for various engineering topics and only used the classes as entertainment.
I think the first semester, or at least half of that, you should go to every lecture until you develop taste for what's useful. By the time you do your master's you should have a good sense for what classes are serious and have useful lectures and where you can just cram last minute because it's a bullshit class and where you're better off studying from a textbook.
do you have any evidence that CalTech's pedagogy is immune to today's threats and trends in education everywhere else? (no)
Nope.
When I was there, a long time ago, exams were timed and were usually open book open note. Blue books were filled in. You were trusted to adhere by those rules, and most students did their exams in their dorm rooms.
The evidence that the students honored the rules was some exams resulted in a 50% failure rate.
As for me, I went there because I wanted to learn the material. I did not care about getting a diploma. (Mine is in the basement somewhere.) I did not take any "easy A" classes, because I wanted a return on my time and tuition investment. (Though, easy A classes were hard to find at Caltech.) I wasn't even going to attend graduation, but my parents showed up and I attended to please them.
The classes, year by year, were dependent on mastering the previous year's classes. So if you cheat with AI, you're digging yourself into a bigger and bigger hole. Caltech rewires your brain. If you don't learn the stuff, you're going to be one of those EEs who carries around a card with V=A*R, V/A=R, V/R=A printed on it.
Unlimited retakes would be an enormous amount of work for professors/TAs/teachers.
Optional homework is often a disaster. At best, students would do it right before an exam and the goal of education is not to just pass exams. They’d probably still get a lower score than if they did the homework when they were supposed to.
What I think is better is to have a due date, but just make the maximum 10% each day it is late. So after 2 days, the highest score you could receive would be 80%.
I liked that system because it gave some flexibility with deadlines while still encouraging you to turn things in on time.
What you described about exams is like how I approach most certification exams: go through a test bank of questions until I get 80% then take the real test.
The study was done in China, where making the class easier to improve pass rates isn't really a thing, as the Gaokao operates more like a stack ranking where it doesn't matter much how well you did in an absolute sense, only that not too many others who did better than you are competing for the same spot. Unlimited retakes are possible in theory, except each costs you a year of your life and except for some extreme cases of repeat test-takers, most people are going to give up after just one or two bad results. Homework is definitely not optional, but going home from school is. (Due to the hukou system, children often have to go to school far from where their parents work, so boarding schools with teacher-supervised self-study are common.) I don't see how holding back for individual classes is supposed to work considering scheduling conflicts.
> F students will realize they have to study, start actually learning and then pass on the second time
Do you have evidence for this beyond your friends (who were accepted into college)?
> If you try and force kids to study when they don't want to "for their own good" you're not actually helping them. Everyone needs to find their own path. Let people fail.
this rhetoric is pretending to be an alternative to coercion. IMO the ideas you are talking about are well trodden and are still coercion nonetheless.
I could be mistaken but my interpretation is that "studying for your own good" is being contrasted with "studying so you don't fail". If you actually fail students, they'll have a second motivation beyond trusting some authority figure's advice. Namely consequences.
Coercion explains it perfectly. A lot of kids see no future because of the environment they live in, so they have exactly zero motivation to study
You are missing the point. Getting a certificate for all of the programs you passed and still graduating could completely change the course of a person’s life. Community college has certificates and associates degrees to let people start working when they reach a practical limit in their general education.
It's all cute but naively glances over what this all boils to - competition.
Education system, contrary to popular belief/name, isn't tailored to educate but to select winners and losers which then will be picked on the job market.
That's why it seems absurd when you think about it as an institution that aims to educate. That's because that isn't the real purpose of it. The purpose is to stratify and classify early.
Schools do not operate in a vacuum; they serve as credentialing gatekeepers for a hyper-competitive capitalist job market. If everyone could easily retake exams until they got an A, grades would lose their primary utility for employers and universities: differentiation. Society relies on schools to provide a neat hierarchy of candidates that for one reason or another thrived in difficult environment of adolescent schooling.
The system often prioritizes compliance, endurance of boredom, and social maneuvering over actual critical thinking precisely because those traits align with corporate hierarchies.
Moneyshot: "The results are eye-opening. After six months, pupils using ai saw their average homework score rise by 18% across all subjects. The time they took to complete each assignment fell from an average of 64 minutes to 45. But come exam time, the same students scored 20% below their classmates who had not called on ai’s help. Homework scores once predicted exam performance; now those who score highest are, perversely, more likely to do worse in exams."
Study design: "David Stromberg of Stockholm University and Victor Lei and Wu Yanhui of the University of Hong Kong set out to fill the gap. They tracked 27,000 pupils aged 12-18 in China, where ai adoption has been fast. Around 80% reported using models such as Doubao and DeepSeek; the other 20% formed the control group."
> Homework scores once predicted exam performance
As with most training the journey is the point, not the destination.
That said, I think smart use of AI could help. It could explain concepts in a way that might help you understand better, it could probe your knowledge in a more dynamic way by tailoring questions, and so on. This requires the AI be restrained by some harness, not free to write down the answers for you.
> help. It could explain concepts in a way that might help you understand better
People say that all the time, but does anyone really think a lack of good explanations for things is a limiting factor in 2026? Or even 2010?
When I was in college, we had office hours with the prof and the TAs, we had group study sessions, and private tutors. Or you could ask your friends. All of those required going somewhere at a specific place and time and asking someone else to give up their time for you.
AI gives you a way to get the same help without asking another human. Say of that what you will, but not everyone was comfortable asking other humans for help even back then.
Back when I was studying there were topics which I had a really hard time grasping. Having a personal tutor that I could ask specific questions and have a back and forth with would have helped a lot I think. At least the few times I did have such an opportunity that was definitely the case.
Kids are, by and large, lazy however. Hell, adults are lazy, this isn't even an indictment on kids or even people in general, I think our bodies & brains are simply wired to seek the path of least resistance from an evolutionary POV.
Kids won't be using LLM tooling as a personal tutor following some sort of Socratic method, they'll ask it to solve their home/coursework for them and blindly copy/paste the answer. Hell, they'll just manually copy down what's on their screen if copy/pasting isn't possible for whatever reason.
Obviously exceptions exist, but I'd wager from being an ex-kid myself the type of kid who would genuinely use these tools for actual proper self-tutoring would be an extreme rarity.
> This requires the AI be restrained by some harness
How would this magic harness look like and why would anyone use it?
"I need help with my homework, but I don't want answers, just guidance on how to get to the answer".
Put that in your agents.md.
I'm way beyond homework, but I do appreciate the compliment on seeming young!
I meant: what's going to stop the average lazy student from ignoring whatever harness the university recommends and instead use an unrestricted LLM, thus learning nothing?
> what's going to stop the average lazy student from ignoring whatever harness the university recommends and instead use an unrestricted LLM, thus learning nothing?
Nothing until they start failing exams and maybe seeing the error of their ways. My suggestions wasn't for you, it was for the smart student who wants to use AI to enhance their learning but not have it do all the work.
But cheating isn't new. People have been cheating in school for centuries. It's easier now, but the consequences were always the same. At some point the chickens come home to roost and you pay the piper.
> But cheating isn't new. People have been cheating in school for centuries. It's easier now, but the consequences were always the same. At some point the chickens come home to roost and you pay the piper.
My concern is that with LLMs and above average bullshitting abilities, the chickens might never come home to roost.
Taking a forklift to the gym increases the number on the bar, but mysteriously your gains go down.
To extend this analogy:
Using the forklift as a spotter and to assist in loading weights increased gains. Then again, a human can do all those things, and provide real human connection.
I am waiting for AI Viva Voce (“AVV”). I am really surprised we aren’t hearing more about this idea?
Try it for yourself by making a prompt like this (adapt as needed):
Pretend I am an undergraduate student of Computer Science. I am learning about early microprocessors from the 1970s. I want you to ask me an examination question as if you were doing a viva voce exam with me, to test my understanding of concepts. I want you to receive my answer and then based on what I said I want you to ask me a more specific question to probe my understanding. Repeat this interaction up to 5 times. Then grade my understanding so far, by giving me a pass, merit, credit, or distinction. Can you explain how you arrive at the grade based on my answers and your expectation of undergraduate knowledge of microprocessor theory?
I've tried stuff like this, but the problem is that you don't know if it's answers are correct, and more importantly, you don't know if its questions are on base or not. Plus its too easy to go off on tangents, especially when you ask it questions to clarify. You can get some more success by writing down a framework beforehand, but openended dialog is still not a good way to learn with AI, from my experience.
I mean, everyone I know has been using socratic dialogue for months to aid understanding, this seems like a variation of that. The issue is that if the answer is too easy to obtain it's hard to have the discipline not to cheat.
If you structure learning in such a way that makes learning just a means to some end, and overindex on that end being the ultimate goal, that's what you get.
This is a pedagogical problem that AI merely exposed. Educators need to figure out How to make students choose the scenic route instead of having them optimize for the most efficient completion of a task.
School has always been about creating compliant workers, not educating people. They will double down on "performing the right things" and "morals" while the economy will continue on its K shaped path as AI gets more capable.
Could it be that you vastly overgeneralize? While there's no perfect education system, many are much better than the nightmare you describe.
Yea, also the programs are too different. My psychology bachelor felt like a tea party. My computer science master was 60 hours per week during the hardest courses (which were supposed to go for 20 hours per week - so I could only do one course if it was at this level). My game design master felt more like we were dumped into an art school masquerading as a psychology/CS master but it really was art school. My information science bachelor felt like the only "normal" study program.
All of this was at university of which 3 of them were at the same university.
Especially the artsy game design program definitely did not feel like it was preparing me to be a cog in some giant corporate wheel.
If others studies tend toward this result, a question comes to my mind: it is a problem with AI or our educational system? or both? I would like to have some comments...
Glad to have scientific results on this, though in my view you could get this from first principles. Homework is onerous but it forces you to learn the material and get it into a configuration that works in your head, which you then validate with the exam.
as a fun aside here from the abstract it does come down to how you use it:
> AI users who maintain similar homework completion time as non-AI users experience small learning losses.
If the pure completion time of homework goes drops significantly with an LLM tool, I suspect these students are spending the extra time turning the material over in different ways to internalize, which is interesting.
https://archive.is/w1eng
It is stopping some of the adults I work with from learning.
In the case of some management, is causing them to unlearn, forgetting about proper review and maintenance practices.
I know this study is focused in China but one thing I'd like to understand is the nuance in what kind of student is likely to always reach for ai for homework help as I think (especially in the US, can't speak for other countries) the varying powers that be that can determine the quality and type of education you can receive (does your country have means to buy textbooks per student or shared, do you have funds to give students ipads to take home or no, what sets the bar on how far curriculums can go, are teacher and faculty pay and incentives simply tied to student pass rates, etc.) and those do more in shaping what "use AI responsibly" will look like and why it's so different.
The homework helps the teacher track their performance/progress.
When the signal is removed, this is what you get.
Honestly I wouldn't demonize AI in education. When we were students we were forbidden from using calculators, then the internet and now it's become the norm. It's the same with AI, I think. AI has already become a part of our lives. The important thing is just teaching children how to use it correctly.
Use AI for homework, you will get it done faster and get a higher grade, and then get crushed on the exam.
Teachers have become proctors and not educators.
i’m pretty sure there’s already a correlation between technology in the classroom and worse learning outcomes
I used to skip homework in high school because it was boring and cut into my personal life outside of school (including my job) but I would always do very well on tests.
Then I stopped caring about tests too.
I did the same thing through college. The only reason they passed me and I got a degree is that I built the school’s website and I built personal ecommerce sites for the head of art and his wife to sell their paintings.
I haven’t done shit since like 7th grade.
Worked at Facebook, Apple, Microsoft, same behavior there - did basically nothing for them while making thousands off my games on App Store.
Fuck authority
[delayed]
When you get out into the real world, what will matter is how effectively you use AI. When you are done with school, you can take the "exam" using AI. The problem is that teaching is trying to teach skills that are no longer relevant and is always behind what the latest going on in the industry. Every time without fail you got interns, who after two years at stanford would learn more in the 3 months on a google internship than in those two years at stanford. They learn useless shit.
My own experience is that college is fun. You can spend the time wisely, but school stuff, is not what makes you money.
Make student loans dischargeable in bankruptcy and make colleges under write them. The problem will get solved quickly.
Way before AI, the problem was very similar - universities do not teach practically useful stuff. Its been a permanent complaint about the education system, really, at least for the last 50 years.
AI just gave the students a way to dodge the slog.
But university was never intended to teach the bleeding edge. That would really be impossible in practice. Pre-phd, it is supposed to teach ways to efficiently attack a problem. you can take a bunch of "play courses", and succeeding at any of those requires pretty much only that one skill.
Nowadays, the challenge they face, is to keep teaching "problem attack methods" in a way that can't be trivialized by AI.
Though fundamentally, if you go to university, and evade learning the one thing you can learn there - thats your loss.