AI 时代，斯坦福的学生在学什么？AI Writes the Code. So What Are Stanford Students Learning?

AI has made learning to code cheap. The hard part now is proving a student can actually do it. Only half of that job can be handed to a machine. The other half takes teachers, exam rooms, and time.

In the age of AI, the argument over whether elite universities are still worth it has never stopped.

So, mostly out of curiosity, I talked with friends in the US who are studying CS or doing MBAs, then looked at the place at the center of the AI boom, the Bay Area. I pulled up the course catalogs at Stanford and Berkeley to see what students at the very best universities in the world are actually learning.

Stanford has a rule that is more than a hundred years old: exams are not proctored. The Honor Code, adopted in 1921, says faculty will not take unusual or unreasonable precautions to prevent cheating, and proctoring counted as one of them. Honesty was the responsibility of the students themselves. On April 23, 2026, Stanford's Faculty Senate voted unanimously to allow proctoring of in-person exams. The Undergraduate Senate and the Graduate Student Council voted in favor as well, and the change takes effect this fall.

Cheating and students who don't report it are old problems at Stanford. AI pushed them past the tipping point. Brian Conrad, a professor of mathematics, put it this way: AI is a challenge for colleges everywhere, and the genie is not going back in the bottle. Keith Winstein, an associate professor of computer science, said that because of AI he now sees the merits of proctoring, but finding enough classrooms is a serious practical problem.

The introductory CS courses had already started changing. CS106B added one-on-one, in-person assessments with TAs, with plans to bring them to CS106A. In-person midterms and finals now carry more weight, take-home assignments less. CS106B's own data showed that students who used AI on assignments did worse on exams. Then, in the senior capstones, CS194 and CS210, AI is explicitly allowed. Tight at the start, open at the end.

In 2026, a year in which AI can write a thousand lines of code a day, one of the best CS departments on the planet is calling students back into the room, sitting them down, and having them prove in front of a TA that they know how to do this. Kind of fascinating, right?

Is the school resisting AI? I don't know. I pulled Berkeley's, MIT's, and CMU's fall catalogs and intro course policies too. All four schools are strictest at the same point: assessment. When a piece of homework comes in, can it still prove the student knows how to do it? On September 13, an article about Stanford's course catalog concluded that Stanford is training students to leave behind two kinds of evidence: proof that you can do it yourself, and proof that you can do bigger things with AI.

Then I looked at seventeen business schools. All four CS departments teach both kinds of evidence, but they only vouch for the first. The ones selling the second as a feature are the business schools next door.

01 The scarce thing has moved

For the past twenty years, these departments spent most of their money on taking someone who couldn't code and teaching them to code well, well enough to land a good job. Courses, TAs, homework, projects, office hours: the whole apparatus was built around that one job. And it really was hard, which is why a good CS department was valuable, a good PhD commanded a high salary, and big tech snapped them up the day they graduated.

Since AI arrived, learning to code has gotten much easier for students. AI is online 24 hours a day, and you can ask it the same question twenty times without it getting annoyed. UIUC's CS 124 went ahead and wrote Claude Code into the syllabus as the textbook, which students buy themselves at $20 a month. For a student, the cost of getting a concept explained and getting a first round of feedback has collapsed over the last two years. On the school's side, the cost of courses, faculty, and TAs has not necessarily come down with it.

For the school, the hard problem is now a different one: when a student hands in code that runs, how does the teacher know whether the student can actually do it, or the tool can?

Stanford and Berkeley landed on the same answer: bring people back into the room. It sounds old-fashioned, but every one of these elite CS schools is heading the same way.

The problem is that this approach depends on manpower. Printing exams takes people, proctoring takes people, grading takes people, and an oral exam means one teacher sitting there listening to one student talk for twenty minutes. AI can run the first round of questions for the teacher, but if a student says I got that right and you graded it wrong, someone still has to sit down and listen. Screening got cheaper. Judgment and sign-off did not, and both come out of the same budget.

Most of the unusual moves these four schools made this year point at that gap. Cheating, class size, and honor codes each explain part of it too.

02 The catalog: 70% untouched, new courses stacked on top

That article only counted Stanford.

These four percentages can't be compared side by side. There is no agreed standard for what counts as an AI course or which research course numbers to drop, and a different method can move the count by several courses. CMU's 18% is especially low: most of its AI teaching lives in the Machine Learning Department, the Language Technologies Institute, and the Robotics Institute, none of which are in this table.

Not one core course disappeared at any of the four schools, and together they still make up more than 70% of the catalog. Stanford's fall schedule still has the backbone courses, CS103, CS106A, CS106B, CS107. Two things changed: old courses were reworked from the inside, and new courses were stacked on top.

The clearest example of an old course is Berkeley's CS 61A, the biggest intro programming course on campus. The Fall 2026 version adds a two-week agent unit, and the final project requires students to build a large piece of software with a coding agent. Everything else in the foundations stays, and the syllabus says most coursework still requires students to write correct programs by hand. On regular homework you may not let Copilot or ChatGPT write your code; on the final project you must use an agent. Same course: banned at the start, mandatory at the end.

The intro course isn't losing students either. CS 61A enrollment rose from 1,269 in Fall 2025 to 1,469 in Fall 2026, while nationally, enrollment in four-year undergraduate CS programs fell 8.4% year over year.

The new courses split into two tracks: undergraduates learn how to use agents, graduate students learn how to build them.

All seven gates have the same shape: set a bar first, then let AI in. CS329Z bans frameworks, CS 124 allows agents only on the final project, 6.S950 locks the door behind three prerequisites, and CS 61A writes the order into the syllabus. Put plainly, these courses teach how much you have to know before AI gets its turn. Get the order wrong and the courses lose their reason to exist.

While schools are still debating when to allow AI, work inside companies already looks completely different. Karpathy, who once taught CS231n at Stanford, said on the No Priors podcast in March 2026 that around December 2025 his split between writing code himself and handing it to agents flipped from 80/20 to 20/80, and that he has barely typed a line of code since. At Notion, every pull request is now written by an agent, and every one is still reviewed by a person. Writing can go to the machine. Review still takes a human, and to review it you have to be able to write it.

03 Assessment: after homework stopped being trustworthy

Teaching itself didn't change all that much this year. How students are tested changed the most. The intro course policies at the four schools fall roughly into three groups.

Group one: ban it outright. MIT 6.101 doesn't even let AI explain concepts; the syllabus puts conceptual help off limits too. CMU 15-122 has students turn off AI extensions in their editors. Princeton's COS 126, Cornell's CS 1110, and Georgia Tech's CS 1331 are in the same camp.

Group two: you can use it, but you have to be able to explain your work in person. The CMU 15-112 syllabus: AI is allowed, even encouraged, but the only way to earn points is to fully understand your code and be able to explain it to an instructor or TA. Its last line: You cannot pass the course if you over-rely on AI. It's that simple. MIT 6.100A allows AI only as a tutor, and Harvard's CS50 allows only its own rubber duck bot, the CS50 Duck. Stanford CS106B's one-on-one TA assessments test the same thing.

Group three: ban it on homework, require an agent on the final project. That is Berkeley's CS 61A. To ban and require at once, the syllabus sets up three lines of defense: an editor extension called Provenance records your entire process of writing an assignment and gets submitted with it; after you turn it in, an AI called Preceptor questions you about your code, and if you can't answer you can book a human instead; and of four exams, three are back on paper. So Berkeley is now using one AI to interrogate you about whether you secretly used another AI. In effect it admits something the other schools won't: a ban can no longer keep students away from AI, so the course can only make sure they can still answer for the work afterward.

A large part of the difference between these three approaches is how much manpower a school is willing to put on the job. A ban is the easiest policy to issue and the hardest to enforce. Explaining in person means someone has to sit there and listen. Group three looks like the most labor-hungry, so it runs several mechanisms at once. I haven't found comparable public figures on TA headcount or student-faculty ratios across the four schools, so this can only be inferred from how each approach works.

This year agents joined the banned list. Berkeley's CS 61B data structures policy says no large language model tool may read, analyze, or interface with your codebase, and it names Claude Code, Codex, GitHub Copilot, and Cursor. MIT 6.102 and CMU 15-213 word it differently and mean the same thing. From punch cards to compilers, from Stack Overflow to agents, the tools for writing code have turned over several times, and the evidence these four top CS departments want has never changed: you have to be able to do it yourself.

On the hiring side, plenty of companies already allow AI in interviews. Karat surveyed 400 engineering leaders in the US, India, and China in January 2026: 38% of US companies let candidates use AI in technical interviews, and 68% in China. David Singleton, the former CTO of Stripe, described on the Latent Space podcast in March 2026 how his new company, Dreamer, hires engineers: first a hand-coding screen to make sure the fundamentals are there, then the candidate builds a complete product with Codex or Claude Code while the team watches how they direct the agents. That is almost exactly how Stanford and Berkeley sequence it: write by hand first, open up the tools at the end.

Schools are policing agents in the exam room and companies are allowing AI in the interview, but both want one look first: with the tools taken away, what is left in this person's head?

The cost is already showing up in the numbers. In Spring 2026, Berkeley's CS 10 had a failure rate of 35.3%, against no more than 10% in earlier years; Dan Garcia, who teaches it, told the campus paper that nearly 30 students were caught using large language models to cheat on take-home exams. CMU's 15-112 had a record number of drops in Spring 2025, and one student who came for help described their AI use as an addiction.

Harvard and Stanford ran into the same thing: once AI help went live, people stopped coming to human office hours. CS50 responded by adding three proctored in-class quizzes, cutting the weight of homework from 60% to 35%, and adding a ten-minute in-person conversation with a TA. At Stanford, turnout at intro course help hours has been falling too. MIT, for its part, issued a university-wide report in August 2026 recommending oral exams, portfolios, and in-class conversations as forms of assessment.

Most of the moves on the public record started after a problem had already appeared, so this round of assessment reform feels mostly like damage control. It is too early to say any pedagogical camp has won.

The schools are the ones changing the rules. The people paying for the changes, through penalties, missed degrees, and bad habits, are the students who happen to be enrolled right now.

04 Teaching is short of people, research is short of compute

Stanford and Berkeley are both calling students back into the exam room, but the price each one pays is completely different.

Stanford is rich: its endowment was $40.8 billion in fiscal 2025, and its constraint is the one Winstein named, finding enough rooms. Berkeley is a public university. Its number of CS graduates was 1,029 in 2024-25 and is projected at about 350 in 2026-27, down by two thirds. How many of those students moved to EECS, data science, or other majors, Berkeley hasn't said. Department chair Jelani Nelson told the campus paper in April 2026 that student demand was not behind the shrinkage and that the cause was the high cost of teaching; over the same period, enrollment in the intro course CS 61A kept rising.

The cost of teaching sounds like an excuse. But for fiscal 2025-26, California's budget cut $271 million from the UC system, and in March 2025 the UC system froze hiring, which Berkeley began enforcing on April 11. Tenured faculty salaries can't really be cut, so a school can most easily move two levers: not opening tenure-track positions, and how many students it admits.

To expand computing and data science, Berkeley has had to lean heavily on fundraising. Friends of mine at Berkeley have told me more than once that they get fundraising emails from the school almost every week. The new Gateway building was launched by an anonymous $252 million gift in 2020, and three later gifts totaling about $75 million continued to fund construction, one of which also created two new computing faculty positions.

The new forms of assessment from the last section only widen the gap. Homework can be graded in batches; a ten-minute in-person follow-up has to happen one student at a time. The two schools are stuck in different places: what Stanford lacks is rooms and execution, while Berkeley has already passed the cost of teaching through to the number of students it admits.

For universities, AI has raised two kinds of scarcity at once: teaching needs people to do the verifying, and research needs money and compute to keep up with the frontier.

Research funding has also become much less certain. Stanford's sponsored research revenue was $2.3 billion in fiscal 2025, more than 70% of it from the federal government, a funder facing changes of administration and prone to changing its mind. That money is accounted for separately from undergraduate teaching, though; only indirect cost recovery flows into the university's general funds, so its effect on teaching is not a straight line.

Beyond money, there is compute. Compute, compute, compute: it bears repeating. Stanford's own 2026 AI Index reports that industry produced more than 90% of notable AI models in 2025. Bo Wang, a University of Toronto professor and chief AI scientist at the AI drug discovery company Xaira, says the first question students often ask when joining a lab is how many GPUs it has. On the industry side, Google's Amin Vahdat, who runs AI infrastructure, deals with a different daily reality: in a cluster of 100,000 accelerators, depending on the configuration, something may fail several times an hour. The metric they hold themselves to is goodput, how much useful computation actually gets delivered while things keep breaking. It is hard not to think: it must be nice to be rich.

University labs can't easily reproduce operations at the scale of 100,000 chips. A classroom can only teach the principles; it can't offer hands-on experience at that scale. Academia hasn't stood still, though. Stefano Ermon's lab at Stanford did early work on diffusion models, and DPO started as a rotation project in his group. On the No Priors podcast in September 2026 he said the hard part is that there are never enough resources, and the good part is that academia lets you make contrarian bets. New directions still come out of universities, but the most compute-hungry experiments at scale are increasingly concentrated in industry.

05 The control group: the same tools are a perk at business school

Walk the same tools into a different building and they get the opposite treatment.

Of the seventeen business schools, sixteen have visible moves on AI coursework; CMU Tepper has none we could find. The remaining sixteen are known as the S16, the top business schools in North America. Only two have a standalone required AI course, Harvard and Yale. Tuck, Darden, and Columbia have folded AI into existing required components. Poets and Quants counted 191 AI courses at 20 business schools in April 2026; of the 166 that could be classified, 157 were electives. These are all things the schools announced publicly. What actually happens in class is invisible from the outside.

Harvard Business School's required course is called Data Science and AI for Leaders. It launched in spring 2025 with 930 students. The course comes with a tool called Julius that does data analysis in plain language, no coding needed; 100% of students used it, sending 126,700 messages in a single term. The lead instructor, Karim Lakhani, said: If the 20th century was defined by the MBA wielding Excel, this century will be defined by MBAs working hand-in-hand with AI agents. In the same year, some CS courses were banning agents from codebases while HBS was defining the MBA as someone who works hand in hand with them.

In fall 2025 Wharton launched the first formal AI major in the M7. Its academic policy says the school provides students with ChatGPT, Codex, Claude, and Claude Code. The tools on Berkeley CS 61B's banned list are handed out at Wharton as a perk.

Stanford's own business school, the GSB, has no required AI course but offers 30 AI electives, the most in the M7. Its student AI policy is also the loosest of the seven. The version updated on August 31, 2026 says faculty may not prohibit students from using AI on take-home assignments and exams; they can only decide whether it is allowed in the classroom. At the same university, the CS department is calling students back into the exam room while the business school forbids faculty from banning AI.

The economics run in opposite directions too. Every extra layer of verification in a CS department means scheduling another person's time. For a business school, one more enterprise account is a purchase: once it is negotiated it can be handed out, and it scales far more easily than another oral exam, which is also why it proves nothing. A system of paper exams plus oral exams plus process records takes teachers, rooms, and money that has to be there every year. Any school that wants to copy it needs those first.

Why are business schools and CS departments so far apart? I think the main reason is the cost of the errors each one is answerable for. In engineering, a mistake that slips through ends up in a production system, so a CS department wants to be sure you can take over when the tool fails. Management education asks you to frame problems, allocate resources, and make the call. AI is more like an analytical tool there, and the cost of letting students use it freely is less direct.

So the two degrees are hard in different places. A CS degree is hard because of how much assessment and how many seats a school can staff. An MBA, at its core, still sells the name and the network.

06 The case against

A few people have arguments that cut directly against all this.

The first comes from inside a business school. Columbia professor Daniel Guetta said at an internal seminar that programming gives you a structured way of thinking, and that the trouble with vibe coding is that it throws the baby out with the bathwater.

The second is an experiment. Wharton's Hamsa Bastani ran an experiment with nearly 1,000 high school students in Turkey. The group that practiced with an unrestricted GPT scored 48% higher, then did 17% worse on a later exam without AI. The group that practiced with a GPT tutor that had guardrails performed on par with the control group. Both sides can use this. The unrestricted group fell apart without AI, which says you have to test students without AI to see what they really know. The guardrailed group, which got hints rather than answers, held even with the control group, which at least shows that in this experiment the AI tutor did not weaken how students performed once AI was taken away. But the experiment only tested teaching. It didn't test whether AI could run the first round of questioning in place of a teacher. If it can, the argument above that verification can't be handed to machines at scale would have to change.

Some students are pushing back too. In July 2025, Stanford GSB students publicly criticized their own school, saying AI is devaluing the MBA, that they were learning to write prompts for ChatGPT without writing a line of code, and that several courses had exam averages of 99. An average of 99 doesn't prove students learned nothing, but it does mean grades can hardly do the job of sorting anymore, and more of a degree's ability to set people apart will fall on the school's name.

07 What it means for you

Since AI came to campus, the money has been moving too. AI companies take the new subscription fees; what schools pay is the cost of people to proctor, run oral exams, and follow up. Within what tuition buys, getting explanations and first-round feedback keeps getting cheaper for students, while credible verification and vouching from the school keep getting more expensive. And employers who receive the diploma still test candidates again in the interview.

If you are studying CS now, or have a child who is about to, looking at the course catalog is clearly not enough anymore. Pay more attention to how the courses are assessed: are there closed-book exams, oral exams, does the instructor look at your process records? What your tuition buys is whatever you can still show to prove you know how to do this once AI is taken away. Whether advanced courses put you in front of real codebases, real companies, and real customers matters far more than whether AI appears in a course title.

If you run a company and you're hiring engineers: portfolio projects are losing value fast, and a beautiful GitHub repo might be an afternoon's work. Interviews will look more and more like a thesis defense. Have candidates explain why they made a tradeoff, chase a bug they didn't write, change an unfamiliar piece of code, and say clearly which parts the tool did. Stanford and Berkeley are spending people's time on exactly this now. Students get tested once in school, and you test them again in the interview.

If you are choosing a school, the number of AI courses tells you little. Look at how rigorously it assesses. The share of oral exams, the weight of in-class assessment, intro course capacity, and whether the school has the money to hire more teachers: these four are what I would look at to judge what a degree will be worth in three years.

08 Final thoughts

Stanford kept its unproctored exams for 105 years and can proctor them starting this fall. Berkeley moved exams back to paper. CMU changed how homework works, and MIT stopped at a university-wide recommendation. The same tools that CS departments ban by name are handed to business school students as perks. Only half of verification can go to machines; once a student disputes a result, a teacher's time is still needed.

Looking ahead, to see how AI affects universities, I will still watch Stanford first: how many courses proctoring expands to, and whether CS106B's in-person assessments make it into CS106A on schedule. Then Berkeley: whether the number of CS graduates stops falling in 2027-28, and what stops it, fundraising, higher tuition, or lower verification standards.

Stanford has one more advantage that other schools cannot easily replicate. With the frontier moving into companies, Stanford's answer is to bring the people from those companies into the classroom. AI courses make up 30% of the catalog, and any school can simply offer the same. What's hard to copy is the CS153 speaker list.

The course is called Frontier Systems. It ran in spring 2026, so it isn't in the fall catalog above. Its two lead instructors, Anjney Midha and Michael Abbott, are from the same company, AMP PBC; one spent time at a16z, the other at Kleiner Perkins. The website lists 26 speakers or speaker groups. Sorted by what they do, they form a cross-section of the AI industry:

Two classes a week, graded 65% on attendance and 35% on the final project. A capacity of 500 and still a waitlist; the opening lecture alone has passed 230,000 views on YouTube. The list also includes Microsoft, Norway's sovereign wealth fund, and the US government, none of them local, so the Bay Area can't explain the whole lineup, but it does make showing up much cheaper for a good share of them. The catalog can be copied. The lineup and the radius of the Bay Area can't.

The final project is called The One-Person Frontier Lab: one person plus the right tools, and how far you can scale in ten weeks. I answered that question myself this year. With several agents running at once, somewhere around the seventh, the team didn't fall apart. I did. One day the upstream data stopped matching what it usually looked like, and I couldn't tell whether the source had changed, an interface had broken, or my own program had a bug. The tools can write the code. Noticing that something is wrong is a moment you have to handle yourself. When Stanford calls students back to the exam room to prove in front of a TA that they can do it, that moment is what they are training for.

But once the school has done its proving, will the market accept it? When employers and investors spend money, which kind of proof are they buying? The answer may be neither. Part two is about the market side.

Sources (course and university pages, press coverage, and industry reports)

Stanford assessment: Stanford Daily 2026-03-07 (A new wave of education: From undergraduates to Ph.D.s, how is AI shifting classroom policies?), 2026-04-24 (Faculty Senate authorizes exam proctoring), 2026-05-20 (Stanford to begin exam proctoring following pilot program); Academic Integrity Working Group AY 2025-26 proctoring pilot information (CS103 course site). Course catalogs: cs.stanford.edu fall schedule; eecsis.mit.edu who-teaches-what list; www2.eecs.berkeley.edu fall schedule; csd.cmu.edu fall course list. New courses: agencyai.mit.edu (6.S950); mit-oasys.github.io/ai-engineering (6.S978); cmu-agents.com (11-768); csd.cmu.edu/course/15113/f26 and www.cs.cmu.edu/\~113 (15-113); cs153.stanford.edu; Stanford Daily 2026-04-15; UW fall time schedule and the cse490A2 public repo; www.cs124.org/syllabus (UIUC). Policy texts: py.mit.edu/fall26/info/ai (6.101); www.cs.cmu.edu/\~112/syllabus.html; www.cs.cmu.edu/\~15122/syllabus.shtml; cs61a.org/fa26/syllabus and cs61a.org/cs194-fa26; fa26.datastructur.es/policies (61B); web.mit.edu/6.102/www/sp26; www.cs.cmu.edu/\~213/academicintegrity.html; cs50.harvard.edu/college/2026/fall/syllabus. Assessment and its costs: Daily Californian 2026-06-03 and 2026-04-16; The Tartan 2026-02-09; Harvard Crimson 2026-09-03; aiandeducation.mit.edu/report (2026-08-13). Finances: California Legislative Analyst's Office, The 2026-27 Budget: University of California; UC hiring freeze and Berkeley's implementation at evcp.berkeley.edu, March to April 2025; Berkeley computing and data science fundraising and the new building at cdss.berkeley.edu (Gateway groundbreaking brings new opportunity for computing, data science) and inspire.berkeley.edu (Three gifts transform computing and data science at Berkeley); NSF NCSES Higher Education Research and Development Survey FY2024; facts.stanford.edu on sponsored research revenue and the federal share; NACUBO-Commonfund Study of Endowments 2025 (Stanford FY2025). Enrollment: National Student Clearinghouse 2026-08-11. Business schools: Poets and Quants 2026-04-28, 2026-04-22, 2025-07-24; hbs.edu/news 2025-03-28; tlhub.stanford.edu 2026-08-31; news.wharton.upenn.edu 2025-04-02 and the mba-inside.wharton.upenn.edu academic policy; business.columbia.edu/insights; mitsloanedtech.mit.edu/tools/claude; kellogg.northwestern.edu 2025-05-30; som.yale.edu 2026-05-26. Industry: karat.com 2026-01-07 (AI Use in Technical Interviews, 400 engineering leaders); hai.stanford.edu 2026 AI Index, research and development chapter; No Priors podcast episodes with Andrej Karpathy (2026-03-20), Notion's Simon Last (2026-03-12), and Stefano Ermon (2026-09-18); Latent Space podcast episodes with David Singleton (2026-03-20) and Xaira's Bo Wang (2026-07-21); Sequoia Training Data podcast with Amin Vahdat (2026-10-06). The case against: knowledge.wharton.upenn.edu (Bastani experiment); coverage of a Columbia internal seminar.