Clara Bennett: Long week — but honestly it was this one thing I read that I haven't been able to shake. A math professor at Caltech, seventeen years in, quietly stopped telling students whether their answers were right or wrong.
Finn Brooks: Wait — stopped grading for correctness? Like altogether?
Clara Bennett: Altogether. Yi Ni now awards credit for completion only. Business Insider broke it November 16th. The proctored exams are still there — that's still the competency gate — but the homework itself? Turned in means credit. Right or wrong doesn't enter into it.
Finn Brooks: No but — okay, so that's not a small tweak to a syllabus, that's like... restructuring what the whole homework ritual is even doing.
Clara Bennett: Exactly — and that's what we're digging into today. What is homework actually for? Because Caltech has no university-wide AI policy; their Center for Teaching, Learning, and Outreach points instructors to set their own course-level rules. So Yi Ni made this decision completely alone, inside one introductory math course, because he couldn't distinguish student work from AI output anymore. And once you can't — now, what does the grade even mean?
Finn Brooks: It means nothing, right? Like that's the whole problem — it's a participation sticker at that point.
Clara Bennett: In practice, yes. Though I'd push back slightly on framing it as the problem. It might be the honest answer to a problem that was always there.
Finn Brooks: Wait, was always there? Like you're saying homework was already broken before AI showed up?
Clara Bennett: Think of it like a receipt. Homework was supposed to be two things at once — a low-stakes practice tool, and proof you'd learned something. But a receipt only tells you someone went to the store. It doesn't tell you they know how to cook.
Finn Brooks: Ohhhh — okay, so AI just made it trivially easy to fake the receipt.
Clara Bennett: Exactly that. And the cooking knowledge — the actual understanding — was never verified by the receipt in the first place. Now, Richard Borcherds at Berkeley, he told Newsweek something that I think is the sharper version of this. He said AI can now solve many undergraduate math problems — better than the students themselves.
Finn Brooks: Better. Not just as good — better than the students.
Clara Bennett: Right — and that's not 'cheating got easier.' That's the tool being more competent than the learner at the core task. MIT's Meng-Chi Ed Chen framed it even more sharply: traditional homework now measures, at best, a student's ability to prompt AI effectively. That's the thing it's actually grading.
Finn Brooks: So has this whole time... I mean, was Yi Ni basically watching that happen in real time? Like he sees the submissions, he can't tell anymore, so he just — stopped pretending the signal was real?
Clara Bennett: That's exactly the read. But here's what complicates it — one of his own students, a sophomore named Sathvik Redrouthu, told Business Insider he deliberately didn't use AI on the homework. Because he wanted to actually understand the material. So the system now can't distinguish that student from anyone who just copy-pasted a solution.
Finn Brooks: So the student who did the work and the student who faked the receipt look identical to the grade. That's — yeah, the receipt analogy just completely holds.
Clara Bennett: And that's exactly where I want to push — because it means the fix can't just be 'detect cheating better.' So what does actually work?
Finn Brooks: And this is the part that actually got me — the answer exists. Like, we know what works. Phil Gressman at Penn replaced his final written exam with individual oral exams back in 2020, kept the format every year since, and his whole thing is that the follow-up question is the tell. You can't fake understanding when someone's asking you why in real time.
Clara Bennett: Right — and there's a version of that at scale too, isn't there?
Finn Brooks: University of Michigan, STATS 413, fall 2025. Picture a Thursday afternoon — three students, one TA, ten minutes, no laptop. They replaced the homework lab entirely because AI could just one-shot the whole thing. Now it's small-group tutorials, talking through problems out loud. That's the pilot. That's the actual replacement.
Clara Bennett: So we have Phil Gressman at Penn doing it since 2020, Michigan running a live pilot in fall 2025 — that's not theoretical.
Finn Brooks: No, it's not theoretical, that's the thing — but then the research just immediately turns around and says these approaches are, quote, 'impractical for large-enrollment courses without major infrastructure investment.' Like, in the same breath. SUNY FACT2 built out this whole dual-lane framework — one lane for in-person oral viva voces, one lane for graded human-AI collaboration — and it's genuinely the most structured policy attempt I've seen, but it still just... assumes the infrastructure exists.
Clara Bennett: So we know what works and we're mostly not doing it.
Finn Brooks: We're endorsing it in theory and hoping someone else pays. A 300-person intro stats course can't do ten-minute individual orals — the TA hours alone would be, I mean, the numbers just don't work without fundamentally rehiring how universities staff assessment.
Clara Bennett: In practice, that's the wall. The method is proven; the institution isn't structured to run it.
Finn Brooks: And honestly, that resource wall gets even stranger when you factor in that the rules aren't even consistent across courses — that part gets into something we're going to get to, but the short version is a student trying to figure out what AI use is even allowed is navigating five different answers depending on which classroom they walked into that morning.
Clara Bennett: And that navigation problem — five courses, five sets of rules — that's not hypothetical. Caltech has no university-wide AI policy. What 'unauthorized AI use' means literally changes when a student crosses the hallway.
Finn Brooks: And it's not just Caltech doing that — this is the survey number that stopped me cold. Someone looked at roughly 146 U.S. university AI policies, and only 26% have rules written specifically for theses or dissertations. The capstone work. The thing you're supposed to point to and say, 'that's what I know.'
Clara Bennett: And 8% leave it entirely to instructor discretion, right? So almost one in ten institutions has essentially said — we haven't decided.
Finn Brooks: Fully haven't decided. And then on the other end you've got Stanford — their 2023 Office of Community Standards guidance treats AI like getting substantial help from another person. Bars it from completing assignments. That is a clear institutional stance. And then there's Skidmore, which prohibits AI unless explicitly authorized and actually deploys Turnitin to catch violations — but the academic consensus is that detection tools are, I mean, they're just not reliable enough to enforce that.
Clara Bennett: So the institution with the strictest posture is enforcing it with a tool that doesn't reliably work.
Finn Brooks: Which means — wait, that's actually the worst of both worlds. You've signaled prohibition, students know detection is inconsistent, and the students who follow the rule anyway are just... absorbing a disadvantage the rule can't actually enforce.
Clara Bennett: Now, the equity layer here — that's what I want to name directly. Students at well-resourced institutions get Phil Gressman's oral exam redesign or Michigan's STATS 413 tutorial pilot. Students elsewhere get an ambiguous policy and a Turnitin flag that might be wrong. The credential says the same thing on both diplomas.
Finn Brooks: And the diploma is quietly — I mean, what is it even certifying at that point?
Clara Bennett: That's the question the degree hasn't caught up to yet. The credential looks identical from the outside — and the gap between what it promises and what it verifies keeps widening every semester institutions don't decide.
Finn Brooks: That gap keeps widening. That's the thing that settled into me. And it circles back to Yi Ni, right — seventeen years in, one introductory course, one quiet decision. He didn't declare a crisis. He just... stopped pretending the signal was real.
Clara Bennett: And that's where I actually land. He didn't break the homework. The homework was already not doing what everyone agreed it was doing — AI just made the fiction too visible to maintain. The honest move was to stop grading for accuracy, not because accuracy stopped mattering, but because he couldn't verify it anymore.
Finn Brooks: Which leaves the degree itself holding that question. If the exam is the only real gate now — the only place independent understanding actually gets tested — then what does the credential certify when institutions still haven't decided whether that gate is enough?
Clara Bennett: I don't think we answer that today. I'm not sure the institutions have either.
Finn Brooks: No, they haven't. Good conversation — genuinely.