A professor at Brown University gave his students a take-home midterm exam. 86 students took it. The class average was 96 out of 100, which is high and surprising until I tell you that 40 of them scored a perfect 100, which is just insane. For context, the historical average on this specific exam across nearly two decades of teaching it ranged between 65 and 80. And this version of the exam was deliberately harder than anything he had said before. Gosh, I wonder what they had at home that made such a big difference. The professor suspected something was wrong, so he told his students the final would be in person. Immediately, 27 of them dropped the course. Nine more simply didn't show up, which seems more like an evacuation than a drop off period. On the take-home, the class answered it beautifully. In person, the average score on that same question was 10%. Which only means if you handed the exam to a room full of exceptionally bright golden retrievers and they randomly guessed by chewing on the Scranton, they statistically would have performed at a similar cognitive tier. Y Combinator co-founder Paul Graham posted the scatter plot on X. It went viral. Two Google DeepMind staffers weighed in as well. The professor, Roberto Serrano, who has taught welfare economics at Brown for 34 years, said he received hundreds of emails and was overwhelmed by the response. The story most outlets are running right now is pretty straightforward. Students cheated with AI, professor caught them, scatterplot went viral, and all of that is true, but the cheating is a symptom. 40 students scored 100 on material they couldn't answer 6 weeks later. They fed a take-home exam to Chad GBT, collected their grade, and walked away with nothing in their heads.

Professor Serrano's course is called welfare economics and social choice theory. It's essentially advanced mathematical economics. We're talking formal proofs, rigorous logical argumentation, the kind of material that requires you to actually think. The kind of course that sounds less like a degree requirement and more like a psychological experiment designed to see how long a human brain can endure pure Greek letters before it simply surrenders. In previous years, the course attracted between 8 and 30 students, which is honestly not a lot. It's difficult. It's niche. It's not the sort of thing you take for an easy grade. But this spring 86 student enrolled. Serrano attributes the spike to one thing. He had announced the exams would we take home. He made that decision for compassionate reasons because in December 2025, a gunman entered an academic building at Brown and unfortunately killed two students, injuring nine others. Many students expressed anxiety about sitting in classrooms. The take-home accommodation was, as SA put it, appropriate. I agree with that. Well, he also designed the midterm to be harder than usual, reasoning that unlimited time would allow students to engage more deeply with challenging material. Instead, he got 40 perfect scores and a class average of 96. When he and his graders ran the exam questions through Chad GPT, the AI produced answers that were, in Serrano's words, kind of correct, but very off and with a very convoluted style. One question could most obviously be solved using a direct argument. Chad GBT and many of his students used a contradiction argument instead. It arrived at the right answer, but through a route Serrano described as very contrived and clearly not produced by human reasoning. It's essentially the classic AI signature, taking the scenic route through a burning building just to hand you a perfectly dry cup of water. So Sana told his class he believed there had been massive cheating. He gave them a chance to prove him wrong. If the grade distribution on the in-person final looked similar to the midterm, he would let the midterm stand. If it didn't, which is of course what he expected, he'd void the midterm entirely and reweigh the final. Which is the professor version of basically saying, "All right, the vibes are off. Everyone back in the cage." The results were catastrophic, and I'm sure you can guess in which direction. The average dropped by nearly 50%. 19 students failed the course, three scored zero, and the students who dropped the course the moment they heard in person, well, most of them had scored 100 on the midterm. Brown's associate dean for the academic code, Love Wallace, described academic code violations as almost never malicious and suggested they often come from a split-second decision. Serrano was not impressed. "You cannot seriously tell me this was a split-second decision," he said. Many of those students cheated all the way to 100. A split-second decision is accidentally grabbing the wrong oat milk at the grocery store. Copying and pasting an entire mathematical proof line by line while ignoring the fact that your chatbot might be hallucinating a dialect of Latin is not a split-second decision. That kind of sounds more like a fulllength feature film of premeditated academic fraud.

Now before I get into the broader implications, I want to pause on something that I haven't seen anyone discuss so far. What was this exam actually asking students to do? Fundamentally, it was asking them to reproduce mathematical proofs to take a theorem, something that has already been proved often centuries ago, and demonstrate that they could reconstruct the proof from scratch. We're essentially asking 20-year-olds to perform intellectual cosplay. We want them to dress up as 18th century French mathematicians and recreate the wheel purely to prove they know that the wheel is round. And I want to be honest with you here. As somebody who holds a PhD in computer science and who spent four semesters being personally victimized by probability theory and God knows only what other branch of mathematics, I have complicated feelings about this. Look, somebody already proved that oxygen exists. Antoine Lavoier figured it out in the 70s7s and humanity has been contentedly breathing ever since without requiring each new generation to rederive the chemistry. And yet in mathematics the expectation is that students will reproduce proofs of theorems that were established decades or centuries ago as though the knowledge doesn't count unless you've suffered through it yourself. Of course, I'm being semifiticious here. I know the argument for why proofs matter. The proof isn't really about confirming the theorem. It's about training a cognitive muscle. The ability to build a rigorous logical argument with no gaps, no handwaving. No, trust me, bro. And of course, I'm aware that discipline transfers when I'm evaluating model outputs that look plausible but are subtly wrong. I'm doing a version of the same thing. Finding where the logic breaks, identifying the hidden assumption. The proof is the gym. The theorem is the weight. But there is a difference between going to the gym to build muscle and being forced to move a pile of bricks from one side of the room to the other just because the supervisor enjoys the sound of heavy breathing. Fundamentally, there is a question of diminishing returns that I think gets lost. I studied graph theory, algorithms, and computational complexity across multiple semesters. I cannot reproduce most of those formal proofs today, but I know what a minimum spanning tree is. I know when a problem is NP hard and I should just stop trying to find an exact solution. I know where to look when I need a specific algorithm. I have the map in my head even though I've forgotten the street names. I know how to navigate the continent. I just can't tell you the exact placement of every specific pebble on the highway. And frankly, if I ever need to know the pebble's layout, I'm just going to look it up, not cry until the ghost of Pythagoras whispers the answer to me. Would half the coursework have built 80% of that map? Would a shorter program plus guided tools assisted exploration have built 70? I genuinely don't know. And to my knowledge, nobody does because nobody's running that experiment at scale. And this matters because Serrano's exam, the one that produced the scatter plot now circulating around the internet, was testing whether students could reproduce a proof using a specific method that the professor expected. Chad GPT used a different method. It arrived at the correct answer through a contradiction argument instead of a direct one. Sarana could tell it was in human reasoning, but the answer was right. So what exactly was being measured? The students understanding of the theorem or their ability to replicate a specific cognitive route to it. If I ask you for directions to the supermarket and you give me the exact coordinates but arrived there via pogo stick while blindfolded, you still got to the supermarket. I'm not saying the exam was bad, by the way. I'm just saying it was a traditional exam designed to test traditional skills in a world where AI makes those specific skills trivially easy to outsource. And when the cost of outsourcing drops to zero, which is exactly what happened, the exam stops being a test of knowledge and becomes a test of whether you choose to use the tool sitting in your pocket or not.

Now, just how millennials cannot afford houses because we all have Netflix and love avocado and toast. Similarly, students are just always lazy. They're just lying face down on a pile of expensive textbooks, actively refusing enlightenment because the modern youth simply lacks the moral fiber of our ancestors who notoriously love doing advanced welfare economic proofs by candlelight. But we wouldn't be on House of L if we didn't explore the devil's advocate. So, let's give that a go for a second. Brown University has an acceptance rate of 5.5%. Getting in requires years of sustained academic performance. These students are demonstrably capable. So why did the majority of them and the data strongly suggest it was the majority choose to hand a take-home exam to Chad GPT rather than engaged with the material? I don't think they're averse to effort, but I do think they're averse to a specific kind of effort that feels disconnected from what they need. Let's consider the economic context. And I won't go in depth on this because this is not a macroeconomics video much to my shagran but it matters. Brown's total cost of attendance for the 202627 academic year is $97,000. Tuition alone is around $75,000 up nearly 50% in a decade. At that price point, you're not really buying an education. You're buying a very expensive receipt. For nearly a hundred grand a year, that receipt better be printed on 24 karat gold leaf and come with a complimentary house. Total student loan debt in the United States currently stands at $1.83 trillion, spread across 42.8 million borrowers. In Europe, the model is different. Tuition is heavily subsidized or free in most countries. So, the debt burden is lower, but the pressure doesn't disappear. It just shows up elsewhere. Youth unemployment is 15.1% in the EU right now, 16.2% in the UK and 24.2% in Spain. Over a million young people in the UK are currently not in education, employment, or training, which is the highest since 2013. Whether you're drowning in tuition debt or graduating into a job market that just doesn't want you, the result is the same. The credential feels like the only thing that matters and the learning feels like a luxury you cannot afford. Over the last 30 years, the cost of a 4-year degree at a public institution in the United States has doubled, while median family income has increased by just 39% adjusted for inflation. The gap between what education costs and what families can afford has been widening for decades, and students are filling it with debt. When the credential costs this much and the job market looks like that, the incentive structure quietly shifts. The credential becomes the product. The learning becomes optional. And when a tool arrives that lets you get the credential without much of the learning, you're not being irrational by using it. You are just responding to a price signal. If you charge somebody the price of a small suburban home for a passport, just don't be shocked when they use a photocopier to pass through the border. The students who cheated at Brown were not making a moral decision. It seems to me they were making an economic one and the system they're operating in incentivizes exactly that behavior. Add to this that these students are navigating genuine trauma, a campus that lost two students in a classroom 6 months earlier. A workload designed for full-time study alongside the economic pressures that often require part-time work. The cognitive cost of sustained academic effort is real and expecting superhuman output from all humans simultaneously cannot become the standard because it defeats the whole point of what superhuman even is. You can sprint for short periods. But if the system demands sprinting at all times, you don't get heroes, you get burnout and shortcuts. Srano framed it very well. The problem with this technology, he said, is that the cost of cheating has basically gone down to zero. It's very easy for students to succumb to the temptation. He is thinking like an economist and he's right. When you reduce the cost of a behavior to zero and the penalty for getting caught is kind of negligible, the behavior becomes the default. You don't fix that by adding proctors. You fix it by changing what you're asking students to do.

But before I tell you what I think should change, I want to show you what the research actually looks like when you zoom out beyond just one classroom. The scatter plot from Brown went viral because it's visually dramatic. You can see the cheating. It's immediate and legible and makes for an excellent tweet, but it's also one course, one professor, 59 students. That's not really good enough for the rigor that we need. So, let's talk about what the research at scale actually shows. In May 2026, Cheerovs, Smeirnov, and Kizelchek published what is currently the largest study of generative AI use by undergraduates. We're talking 95,513 students across 20 major US public research universities published in science, the holy grail, okay? It's the academic equivalent of getting a blessing from the Pope except with a lot more P values and fewer hats. Twothirds of students had used generative AI for coursework. 9% admitted to using it to cheat. And here's the critical finding. Daily users were more than three times as likely to cheat, 26% compared to monthly users at 7%. It's basically a slippery slope type of finding. The more you use it, the more you offload, and the more you offload, the harder it becomes to do the work without it. The study also found that AI adoption and misuse vary dramatically by discipline. Regular use is highest in computer science at 62% followed by mathematics at 53% and business at 51%. The patterns of use are different, the patterns of misuse are different and the skills being tested are different, which means that any response to AI cheating needs to be discipline specific. A blanket policy on this one makes about as much sense as prescribing the same medication for every single illness. The researchers explicitly called for assessment reform at the discipline level and I completely agree with them because it's exactly the same conclusion I reached in my previous video when we talked about Princeton. Which brings me to something that I find genuinely very interesting. This paper 95,000 participants calling for the same structural reform that Serrano's anecdote illustrates has received a fraction of the public attention that the scatterplot got. We respond to stories. We respond to visuals, to YouTube videos. A scatterplot is dramatic, legible, and emotionally immediate. A peer-reviewed study in science is dense, careful, and quietly devastating. It is pages upon pages of data so dry it demands a martini that tastes like the bottom of the Mediterranean Sea. The scatter is essentially a beautifully wrapped piece of academic gossip. It is a visual meme that says, "Look at these dirty cheaters." in a single glance. The science paper is the direct 40-page blueprint explaining that the entire mountain the school is built on is sliding into the ocean. But the blueprint doesn't fit cleanly into a tweet. So, we just keep talking about the gossip. Both matter. Both deserve attention. And the fact that the visual wins tells you something about how we consume evidence as a society, which incidentally is itself a problem of critical thinking.

Now on the question of whether AI is actually damaging students cognitive capacity, the literature is suggestive but not yet conclusive. Gerish 2025 studied 666 participants and found a significant negative correlation between frequent AI use and critical thinking scores mediated by cognitive offloading. An MIT Media Lab study used EEG to monitor brain activity during essay writing and found that participants using Chad GPT displayed the weakest neural connectivity and those patterns persisted even after they stopped using the tool. The researchers called it cognitive debt. Other research tells a completely different story. Alfaria atal 2026 found that standard chat GPT use produced the highest perceived understanding alongside the lowest actual learning. So confidence without competence which coincidentally is also the exact business model of every tech influencer on LinkedIn. However, when the AI interaction was redesigned with what they call pedagogical friction, so Socratic questioning, guided hints, prompts that forced the student to think before receiving the answer, learning gains were significant. A separate study on Socratic AI tutoring found that students reported significantly greater support for critical, independent, and reflective thinking, directly challenging the narrative that AI necessarily makes you worse at thinking. So based on these studies, the conclusion from the literature is not AI is bad for learning. The conclusion instead is it depends entirely on how you use it. Use it as a substitute for thinking and it atrophies the muscle. Use it as a Socratic partner that pushes you to reason and it can strengthen it. The tool is not the problem. The interaction design is. And right now, nobody's teaching students how to use these tools as thinking partners rather than thinking replacements. Brown's own generative AI and teaching and learning committee, whose report was published the same week as the Srano scandal, found that 56% of undergraduates and 67% of graduate students use generative AI daily or weekly. And here's what I find most striking. Large majorities of those students reported being concerned about the impact on their own learning. They are worried about losing cognitive capacity. They know it's a problem and they're doing it anyway. It's like staring down at a third slice of chocolate cake at 1:00 a.m. Your brain is actively screaming, "This is a terrible choice for a long-term structural integrity." While your hand is already moving for the fork, we need rigorous largescale research, actual control trials with significant participant numbers to understand what's happening here. The Cherikov study is a critical major step in the right direction. We need a lot more like this. We need them across disciplines, across countries, across educational models. We are dealing with a technology that is reshaping how an entire generation interacts with knowledge. And the evidence base for making policy decisions about it is still nowhere near where it should be.

But even with what we have, the Brown data tells us something definitely worth sitting with. There are two students in Serrano's data who are worth paying attention to. one scored consistently high on both the midterm and the final with or without AI. They knew the material. Srano said he knew them well and described them as an excellent student. That's the person you hire, right? That's the person every employer should want. Somebody who has actually built the cognitive architecture, who can use AI as a tool because they understand the territory it's operating in, who can look at a model's output and know when it's nonsense. The other students scored consistently low. Poor on the take-home, poor on the final. They didn't cheat. Servano said, I'm quoting, "I admire that person, and I understand the sentiment. Honesty matters. Integrity matters, of course. But I'll be honest with you, admiration for failing honestly is a very strange place for an education system to end up. The goal is not noble failure. The goal is competence. The right response to that student isn't admiration. It's asking how you help them get better." with AI, without AI, with different teaching methods, with different assessment models, whatever it takes. Because right now the system is producing three categories of graduates. People who learned and can prove it, people who didn't learn and can prove it, and people who didn't learn but have a credential that says they did. That third category is the one that should keep every university president awake at night because the market figures it out. It always does. And when employers start discovering that a significant proportion of Ivy League graduates can't actually do the work their degrees claim they can do, the brand degrades whether or not anyone catches the cheating. A degree from Brown is a signal. If the signal stops being reliable, it stops being valuable. Hiring data scientists and technical positions looks very little like a traditional exam these days. You just kind of assume that they can and will use AI. And honestly, why shouldn't they? assessing knowledge should be happening in a similar environment to the same environment they will have when they're actually applying it. Serrano students responded fairly rationally given their circumstances and the institution's response adding proctors avoiding the midterm investigating through the academic code is the educational equivalent of putting a security guard in a building with a structural problem. The roof is skating in. The foundation is turning to dust. But thank God we hired a guy with a clipboard to make sure nobody walks in with an unapproved calculator. The Cherikov study calls for discipline level assessment reform and I agree. But I can't tell you what that looks like for every single field because I don't sit in every field. What I can say is this. AI is not going anywhere. And I don't just mean large language models. I mean the entire spectrum of artificial intelligence, the tools, the systems, the infrastructure. It is embedded. It is expanding. Every scientific discipline is growing, constantly producing more knowledge that needs to be absorbed by people with finite time and finite lifespans. The question of how we educate people in that environment is no longer optional for us. It is urgent. It is not a question that can be answered by tech companies alone or regulators alone or governments alone. It requires all of us. Whatever field you work in, whatever discipline you studied, wherever you sit, please think about it. Think about what skills actually matter in your work versus what you were tested on in school. Think about how AI intersects with your domain. Talk about it here on YouTube, on this channel, with colleagues, with friends, in your own communities. Please stop feeling like AI is something happening to you and start asking what it could look like if it was happening with you. Because on a long enough timeline, and I mean really long here, like 300 years from now, most of us will be dead. Well, except for me because I'm going to be uploaded into the cloud. I'll be hanging out on a server rack in Iceland, thriving on geothermal energy and judging your descendants math skills in real time. But the rest of you will need to have left behind a world where the next 10 generations of humans are equipped to handle what's coming. And that starts with getting education right. Not just proctors or surveillance or scatter plots on social media. I mean the actual difficult discipline specific work of figuring out what it means to learn in a world where the information is always available and the skill that matters is knowing what to do with it. Professor Serrano said we cannot choose to become idiots and he's right. But the choice is not between intelligence and idiocy. It is between an education system designed for the world as it was and one designed for the world as it is. Brown just showed us in one scatterplot how wide the gap between those two things has become.