Last week, Princeton University voted to end an 133-year tradition of unproctored exams. Starting July 1st, every in-person exam at Princeton will have an instructor in the room watching students take it. The faculty vote was a near unanimous. Just one person voted against. And the reason cited was artificial intelligence. Now, if you've seen that headline and thought, well, students were cheating with AI, so the university cracked down. That is the version of the story that most outlets are running. And it's not wrong exactly, but it misses something much more important because this isn't really a story about AI or about cheating or about Princeton. This is a story about an entire model of education that was built for a world that no longer exists. And the fact that one of the most prestigious universities on the planet just tacitly admitted that, even if they don't frame it that way exactly, should give everyone a pause. I'm El. I have a PhD in computer science and I analyze AI developments. To understand what's actually happening beneath this hype, to understand why Princeton's decision matters, you need to understand what they actually had. Back in 1893, Princeton students, the students themselves, petitioned the university to remove proctors from exam rooms. They didn't want to be watched. They believed that a pledge of personal honor was more meaningful than surveillance. And the faculty agreed. For 133 years, Princeton students took exams without anyone looking over their shoulders. They signed a pledge. I pledge my honor that I have not violated the honor code during this examination. And that was supposed to be enough. But it wasn't just a pledge. The system had an enforcement mechanism. Students were expected to report each other. If you saw somebody cheating, you were honor bound to turn them in. The entire architecture depended on peer surveillance. students watching students trust but with a distributed enforcement layer. And here's where the data gets interesting. The Daily Princonian's 2025 senior survey, this is over 500 graduating students, found that 29.9% of respondents admitted to cheating on an assignment or exam during their time at Princeton. Nearly a third. But the number that actually tells the story is this. 44.6% of seniors said they knew about honor code violations. they chose not to report and the number who actually did report a peer was only 0.4%. That is two people out of 500. The systems entire enforcement mechanism relied on students reporting each other. Functionally, nobody did.

Princeton built a justice system where the jurors were also the defendants. And Princeton is not exactly an outlier. at Stanford. Between 2018 and 2020, just two out of 720 honor code violation reports came from students. A chemistry graduate student told the Stanford Daily that cheating had become, and I'm paraphrasing here, part of the fabric of the university. Stanford began introducing proctoring for some exams in 2023 and as of April 2026, has voted across its full governance chain to allow proctoring broadly. At Middbury College, a review committee found that the honor code has ceased to be a meaningful element of learning and living at Middbury for most students. 65% of students self-reported having broken it. When a collegewide proposal to introduce proctoring was put to a student referendum in late 2025, it failed. Only 48.8% voted in favor. Students voted against being watched while simultaneously admitting in large numbers that they cheat. So that is not exactly a policy failure on the university part. That is a culture operating exactly as its incentives predict. They voted to keep the honor code the same way people vote to keep their gym membership aspirationally. So AI didn't break the honor code. The honor code was already a very expensive fiction. What AI did was make it impossible to keep pretending. And here's a number that complicates the narrative even further. According to research aggregated across multiple surveys, in 2012, 17% of students use their phones to text answers during assignments. In 2026, 18% use AI to submit unedited work. The proportion of students submitting entirely machine generated work without any personal engagement has not dramatically increased compared with previous forms of technological cheating. What has changed is the quality of the output. A texted answer from a friend was obviously derivative. An AI generated essay is fluent, structured, and unless the professor has read thousands of similar outputs, indistinguishable from a competent student's work. The cheating rate didn't spike, the cheating ceiling did. AI related academic misconduct now represents 60 to 64% of all cheating cases in higher education globally. Not because AI created more cheaters, but because it absorbed the cheating that used to happen through other channels. Traditional plagiarism cases are declining at almost exactly the rate AI assisted cases are rising. The cheater didn't change, the cheat did. And the detection gap is staggering. Despite a 33% increase in student discipline for AI related misconduct between 2022 and 2026, an estimated 94% of AI generated assignments still go undetected. The detection rate for AI cheating is 6%. For context, that's roughly the same probability as rolling a specific number on a dieice.

Students are essentially gambling and the house odds are spectacularly in their favor. In the UK alone, nearly 7,000 university students were formally caught cheating with AI tools in the 2023 2024 academic year, triple the number from the year before. And those are only the ones that got caught. By the way, the proposal Princeton's faculty voted on makes this explicit. It says that AI tools and small personal devices have changed the external appearance of misconduct during an examination, making cheating much harder for other students to observe and hence to report. In other words, the system relied on cheating being visible to the people sitting next to you. Somebody pulling out a cheat sheet, whispering to a neighbor, looking at somebody else's paper. AI eliminated that visibility. A student glancing at a phone under a desk looks identical to a student checking the time. As the former chair of Princeton's honor committee put it, if the exam is on a laptop, somebody can just flip to another window. The entire enforcement architecture inverted. Cheating used to be visible and socially costly to commit. Now it is invisible and socially costly to report. Princeton's own proposal notes an increase in anonymous reports of suspended violations driven by fears of doxing and social media shaming. Students weren't just unwilling to report, they were afraid to. And here's what I find most striking. It was the students who pushed for proctoring, not just the faculty. The student government surveyed undergraduates and found a majority either favored proctoring or were indifferent. Current and former chairs of the honor committee endorsed the change. The people being washed asked to be watched because the alternative being responsible for policing their own peers in a system where policing was both impossible and socially dangerous was worse. Under the new policy, instructors will be present in exam rooms as a witness to what happens, but are instructed not to interfere with students. Witness is doing a remarkable amount of work in that sentence, given that the entire previous system relied on witnesses who saw everything and reported nothing. Professor Jill Dolan, who served as dean of the college from 2015 to 2024, described the vote as a shame but necessary. And the proposal itself acknowledges that having an instructor supervising examinations will not eradicate cheating. So even Princeton is saying this is not a solution, it's a band-aid.

And that's the part I want to sit with because I think the conversation should be much bigger than proctoring. Here's the question nobody in this debate seems to be asking. What is the exam actually testing for? The traditional exam, you sit in a room, no resources, write down what you know from memory, was designed for a world where knowledge was scarce and hard to access. If you wanted to know about Keynesian economics or organic chemistry or the treaty of Wisfalia, you had to have read specific books, attended specific lectures, and retain that information in your head. Testing recall made sense because recall was the skill. Your ability to reproduce information from memory was a reasonable proxy for whether you'd actually learned. That world doesn't exist anymore. Not because students are lazier or because institutions have failed, but because the fundamental relationship between humans and knowledge has changed. The information is now universally accessible instantly to everyone all the time. The skill that matters is no longer can you store and reproduce information. It is instead do you know what exists? Do you know when to reach for it? And can you evaluate what you get back? I'll give you a personal example. During my undergraduate degree, I took four semesters of probability and statistics. Four, that is two full years of mathematics that I would not wish on anyone I care about. I was personally victimized by Chibbishev Kulmagorov and the entire Russian school of probability theory. Chibbishev, for those who don't know, was a 19th century Russian mathematician who, due to a childhood condition that gave him a limp, couldn't pursue the military career his family intended, spent a great deal of time at his desk and produced theorems of such elegant complexity that students would be suffering through them two centuries later. I acknowledge his extraordinary contribution to mathematics. I also wish he'd gone outside more. He's dead now, so that's kind of nice. And I feel bad for even saying that, but still. But here's the thing that matters about that experience. 10 years later, I could not reproduce a single one of those proofs from memory. Not a single one. If you sat me down right now and asked me to prove trib's inequality, I would definitely fail. But, and this is the critical distinction, I know that it exists. I know what it does. I know when a problem I'm looking at might benefit from it. And because I know that it exists, I know exactly what to look for. I have the map, even if I've forgotten the individual street names. Now, is that knowledge a product of having done the hard work of having sat through those four miserable semesters and worked through the proofs by hand? Honestly, I don't know. Maybe maybe that suffering is exactly what build the cognitive map and there is no shortcut. Or maybe somebody could have taught me the landscape of probability theory, shown me what tools exist, when they apply, how they relate to each other without requiring me to derive every proof from first principle, and I'd have the same map today. I generally don't know the answer to that, but I think it's the question that education should be asking right now and mostly isn't because here's the broader reality.

Human knowledge is expanding constantly. The body of knowledge required for an undergraduate degree in physics today is substantially larger than it was 50 years ago because there is simply more physics. The same is true of medicine. Becoming a reasonably competent doctor already takes the better part of a decade. The same is true of computer science, of biology, of virtually every field. And this only moves in one direction. Becoming a physicist a 100 years ago was not simple, but it was simpler because there was less physics to know. An undergraduate physics curriculum a 100 years from now will contain a century's worth of additional discoveries. The time it takes to become competent keeps growing, but the human lifespan doesn't. There is too much cannon. At some point, the degree becomes longer than the career it's supposed to prepare you for. I want to be clear here that I'm offering this more as a thought experiment than a prescription. I don't have it fully crystallized. And I think the honest thing is to invite you to think about it alongside me rather than pretend that I've solved it all. But I do generally think that as the sum of human knowledge continues to expand, the probability that any single person can meaningfully engage with the entirety of even their own field gets lower and lower. And that changes what education should be optimizing for. The question is what replaces memorization. And I think what replaces it is something closer to the way I actually use my education. Now, knowing the architecture of a field well enough to navigate it, combined with tools that let you access the specifics when you need them. Think of it like a calculator. You know how to add, subtract, multiply, and divide. You have the foundational understanding but you use a calculator for the heavy computation because the skill isn't arithmetic. It's knowing what calculation to run and when. Now extend that principle to education more broadly. An open book exam except all exams become open book because the book is always open. The skill being tested isn't memorization. It's judgment.

This isn't just a theory about education. This is something I actually practice. When I hire data scientists for my team, for example, the assessment looks nothing like a traditional exam. I give candidates a complex scenario, real model outputs, messy data, ambiguous requirements, and I tell them to use whatever tools they want, any of them. I am not testing whether they can write code from memory because in 2026, a large language model can write almost any code. What I am testing is whether they understand what to do with the output. Whether they can look at a model's results and know what question to ask next. Whether they have the cognitive architecture, the map to direct the tools effectively. That requires a fundamentally different skill set than memorization. And it is substantially harder to fake. Anyone can prompt an LLM to write a logistic regression. Not everyone can look at the output and know that it's nonsense. That is what I'm hiring for. If that's actually something you'd be interested in as a standalone video, how AI is changing the way technical hiring actually works, let me know in the comments below. And the data supports this approach as well. A May 2026 Gartner survey of 350 global executives at companies with at least $1 billion in annual revenue found that roughly 80% of organizations deploying AI or autonomous technologies had reduced their workforce. But, and this is the finding that should be on every university president's desk, those workforce reductions had no correlation with return of investment. The companies that cut people and the companies that didn't were seeing essentially the same results. What did correlate with higher ROI was what Gartner calls people amplification. Companies that used AI to make their employees more productive, investing in skills, roles, and operating models that let humans guide and scale the technology. The organizations that improved returns were not those that eliminated the need for people. They were the ones that amplified them.

I actually made an entire video unpacking that Gardner study. What companies are getting wrong about AI and jobs, why the layoffs aren't generating the returns anyone expected, and what the companies that are seeing returns are doing differently. If you haven't seen it, I'd watch that one after this one. But just keep that info in mind and think about what Princeton just did. The university's response to AI was to add a human observer to the room. The equivalent of adding a security guard to a building with a structural problem. Instead of asking, "How do we redesign assessment so that AI becomes a tool students learn to use?" Well, they ask, "How do we stop students from using AI during the test?" One of these questions prepares graduates for the world they're about to enter. The other preserves the format of an exam design in the 19th century. And I want to be very clear. I understand why students sometimes cheat. The pressure is very real. Deadlines, workload, the competition for grades that open doors to graduate schools and employment. I'm not dismissing that. But I do think there is a fundamental truth that gets lost in the mechanics of it all. Education is something you get for you. You are usually paying for it. At Princeton, you're paying quite a lot for it as well. And the purpose of that payment is not a document at the end. It is a skill set. It's the capacity to think, to reason, to know what you don't know and figure out where to find it. And ultimately, this is a decision every student has to make for themselves. What kind of a professional do you want to be? Do you want to get the job because you have the credential, or do you want to be useful in that job? Do you want to progress or do you want to spend your career hoping that nobody notices? Because credentials might get you through the door, but if all you have is the credential and nothing much behind it, you're not going to get very far once you're inside. If you want to be competitive, and I mean generally competitive, in a world where AI can do the baseline work, you need to bring something AI cannot replicate. Good judgment, context, the ability to know what question to ask, and that requires actually having learned something.

I think a person who acquires that skill set without a diploma is in many practical aspects more employable than a person with the diploma who didn't learn anything. Credentials get you through the door. They do not keep you in the room. And institutions like Princeton should recognize that their brand, the thing that makes a Princeton degree means something to an employer, is not the piece of paper. It's the quality of the person holding it. If the assessment model is so broken that a significant proportion of the graduates haven't actually acquired the skill the degree is supposed to represent, the brand degrades whether anyone catches the cheating or not. The market figures it out eventually. It always does. Back in 1893, Princeton students asked not to be watched. They believe that personal honor was a better guarantor of integrity than surveillance. In 2026, their successors asked for their surveillance back, not because they had lost their honor, but because the system designed to protect it had become impossible to operate. The answer to that is not more watchers in the room. It is building something worth trusting. Again, an assessment model that doesn't treat AI as the enemy, but recognizes it as the environment graduates will spend their entire careers operating in. A model that tests not what you can recall under pressure, but what you can do with every tool available to you, the open book exam for the age of the infinite book. Princeton took 133 years to admit the old model was broken. The question now is how long it takes to build the new one. But Princeton isn't the only institution reaching for the wrong lever. Companies are doing the same thing, firing people in the name of AI, and the data says it's not working either. I made a video about what the numbers actually show when companies try to replicate their workforce instead of augmenting it and why the most expensive mistake in AI right now has nothing to do with the technology. That's the one that I would watch next. Thank you all so much for watching. Subscribe and I'll see you on the next