A few weeks ago I came across a handful of stories about how NYU, the University of Chicago, the University of Sydney, and Cal State Maritime Academy are each responding to AI in the classroom. I kept coming back to them, so I wanted to work through what I think is actually interesting here.

None of these schools are still arguing about whether to ban AI. Once a technology delivers real productivity, banning it is close to pointless — students will use it, the industry they’re headed into already runs on it, and a ban mostly teaches students to comply on paper and use it anyway behind closed doors. What these four are doing instead is figuring out where, exactly, to check whether a student actually understands anything.

Sydney runs two tracks in parallel. One assumes students will use AI to produce their coursework and focuses on the practical skills employers actually want. The other requires a timed oral defense after the paper is turned in, where the student has to explain their reasoning on the spot, with no notes and no AI. Fumble that, and the paper wasn’t really theirs to begin with.

Cal State Maritime took a blunter approach. A math professor moved all his lecturing online and spent the freed-up classroom time watching students work problems at the whiteboard, live, step by step. He can’t stop anyone from running their homework through AI at home. But he can pick a room where AI simply isn’t in the picture.

NYU’s business school built something called Viva, an AI examiner that grills students orally after they submit a project. Students can lean on any AI tool they want while doing the work, but the follow-up questions get harder the better they answer — students who’ve been through it say the ceiling keeps rising as long as you keep clearing it. Instead of treating AI as the enemy, NYU turned it into the thing that drags students toward a deeper answer.

Chicago Law went the other way entirely. Its first-year core course is piloting a total ban on laptops and phones, on the theory that wrestling with hard material, unaided, is precisely how professional judgment gets built. This isn’t a new way of testing understanding. It’s an attempt to keep the old one alive by institutional force.

Set the four side by side, and the more interesting question isn’t ban-or-don’t-ban — it’s that each school is answering the same underlying question in a different way: if you can’t keep AI out, where should the real test move? Sydney and NYU are betting on the same thing: that a person’s actual capability still comes down to whether they can function without AI — whether judgment has been internalized deeply enough that, under pressure, someone can reconstruct the whole chain of reasoning alone. In this view, human ability and AI-assisted output are two separate things, and you can test them separately. Cal State Maritime splits the difference, testing thinking as it happens rather than reasoning explained after the fact. Chicago didn’t relocate the test at all — it’s betting the old battleground, unassisted human thought, is still worth defending on its own terms.

None of this happens in a vacuum, either. How a university tests its students is shaped, whether anyone admits it or not, by what employers want. NYU’s Viva model fits corporate taste unusually well, because it’s an outcomes-based test at heart — nobody cares what tools you used, only how hard a problem you actually cracked. I could easily see hiring head the same way: AI-assisted interview tools built to surface whoever can still solve the hardest problems with an AI copilot. That only works under one condition — different candidates using AI still have to produce meaningfully different results. If everyone gets roughly the same output out of the same tool, the whole screening exercise falls apart. What’s interesting is that this probably runs both directions: what employers want shapes how universities test, and what universities produce shapes who employers can actually hire. NYU’s design, in effect, has already pulled that loop forward and installed it inside the classroom.

This connects to something I keep circling back to in my own work on junior engineers. The traditional apprenticeship rung — where a junior joins a company cheap enough, relative to their output, that training them is almost a side effect of hiring them — is thinning out fast. A senior working with AI tools can now often match what a junior alone produces, which raises the bar a junior has to clear just to be worth hiring at all. What companies are quietly dismantling is the implicit subsidy that used to fund a junior’s training. What these universities are up against, I think, is the same fracture from the other side: a university’s traditional value — transmitting knowledge — always rested on knowledge being expensive to obtain in the first place. Once AI pushes that cost toward zero, universities get shoved out of the role of transmitter and into the role of verifier, whether they’re ready for it or not.

My own instinct is that Sydney’s dual-track model is the most honest of the four. It doesn’t pretend to have already solved the problem — it admits there’s no settled answer yet, and lets different students place different bets on which one will pay off. I genuinely don’t know which bet is right: whether operating without AI is just one viable option among several, or whether it’s actually the foundation good AI judgment can’t exist without. But I doubt this stays a question universities get to answer alone. However companies end up training their juniors, however the rest of us end up raising our own kids, everyone runs into some version of this same choice eventually.