[...]

> Take-home exams are dead. Reverting to pen-and-paper exams in the classroom feels like a regression.

Yeah, not sure the conclusion of the article really matches the data.

Students were invited to talk to an AI. They did so, and having done so they expressed a clear preference for written exams - which can be taken under exam conditions to prevent cheating, something universities have hundreds of years of experience doing.

I know some universities started using the square wheel of online assessment during covid and I can see how this octagonal wheel seems good if you've only ever seen a square wheel. But they'd be even better off with a circular wheel, which really doesn't need re-inventing.

I went to school long before LLMs were even a Google Engineer's brianfart for the transformer paper and the way I took exams was already AI proof.

Everything hand written in pen in a proctored gymnasium. No open books. No computers or smart phones, especially ones connected to the internet. Just a department sanctioned calculator for math classes.

I wrote assembly and C++ code by hand, and it was expected to compile. No, I never got a chance to try to compile it myself before submitting it for grading. I had three hours to do the exam. Full stop. If there was a whiff of cheating, you were expelled. Do not pass go. Do not collect $200.

Cohorts for programs with a thousand initial students had less than 10 graduates. This was the norm.

You were expected to learn the gd material. The university thanks you for your donation.

I feel like i'm taking crazy pills when I read things about trying to "adapt" to AI. We already had the solution.

Don't tell me about GenZ. I had oral exams in calculus as undergrad, and our professor was intimidating. I barely passed each time when I got him as examiner, though I did reasonably well when dealing with his assistant. I could normally keep my emotions in check, but not with my professor. Though, maybe in that case the trigger was not just the tone of professor, but the sheer difference in the tone he used normally (very friendly) and at the exam time. It was absolutely unexpected at my first exam, and the repeated exposure to it didn't help. I'd say it was becoming worse with each time. Today I'd overcome such issues easily, I know some techniques today, but I didn't when I was green.

OTOH I wonder, if an AI could have such an effect on me. I can't treat AI as a human being, even if I wanted to, it is just a shitty program. I can curse a compiler refusing to accept a perfectly valid borrow of a value, so I can curse an AI making my life difficult. Mostly I have another emotional issue with AI: I tend to become impatient and even angry at AI for every small mistake it does, but this one I could overcome easily.

In my graduate studies in Germany, most of my courses used oral exams. It's fine, and it's battle-tested.

When I was doing a lot of hiring we offered the option (don’t roast me, it was an alternative they could choose if they wanted) of a take-home problem they could do on their own. It was reasonably short, like the kind of problem an experienced developer could do in 10-15 minutes and then add some polish, documentation, and submit it in under an hour.

Even though I told candidates that we’d discuss their submission as part of the next step, we would still get candidates submitting solutions that seemed entirely foreign to them a day later. This was on the cusp of LLMs being useful, so I think a lot of solutions were coming from people’s friends or copied from something on the internet without much thought.

Now that LLMs are both useful and well known, the temptation to cheat with them is huge. For various reasons I think students and applicants see using LLMs as not-cheating in the same situations where they wouldn’t feel comfortable copying answers from a friend. The idea is that the LLM is an available tool and therefore they should be able to use it. The obvious problem with that argument is that we’re not testing students or applicants on their abilities to use an LLM, we’re using synthetic problems to explore their own skills and communication.

Even some of the hiring managers I know who went all in on allowing LLMs during interviews are changing course now. The LLM-assisted interviewed were just turning into an exercise of how familiar the candidate was with the LLM being used.

I don’t really agree with some of the techniques they’re using in this article, but the problem they’re facing is very real.

Where do we go from there? At some point soon I think this is going to have to come firmly back to real people.

As an aside, I'm surprised oral exams aren't possible at 36 students. I feel like I've taken plenty of courses with more participants and oral exams. But the break even point is probably very different from country to country.

If this is the only way to keep the existing approach working, it feels like the only real solution for education is something radically different, perhaps without assessment at all

No, we do not want to eliminate the pen and paper exam. It works well. We use it.

The oral exam is yet another tool. Not a solution for everything.

In our case, we wanted to ensure that the students who worked on the team project: (a) contributed enough to understand the project, (b) actually understood their own project and did not rely solely on an LLM. (We do allow them to use LLMs, it would be stupid not to.)

The students who did badly in the oral exam were exactly the students who we expected to do badly in the exam, even though they aced their (team) project presentations.

Could we do it in person? Sure, we could schedule personalized interviews for all the 36 students. With two instructors, it would have taken us a couple of days to go through. Not a huge deal. At 100 students and one instructor, we would have a problem doing that.

But the key reason was the following: research has shown that human interviewers are actually worse when they get tired, and that AI is actually better for conducting more standardized and more fair interviews. That result was a major reason for us to trust a final exam on a voice agent.

Reminder: This professor's school costs $90k a year, with over $200k total cost to get an MBA. If that tuition isn't going down because the professor cut corners to do an oral exam of ~35 students for literally less than a dollar each, then this is nothing more than a professor valuing getting to slack off higher than they value your education.

>And here is the delicious part: you can give the whole setup to the students and let them prepare for the exam by practicing it multiple times. Unlike traditional exams, where leaked questions are a disaster, here the questions are generated fresh each time. The more you practice, the better you get. That is... actually how learning is supposed to work.

No, students are supposed to learn the material and have an exam that fairly evaluates this. Anyone who has spent time on those old terrible online physics coursework sites like Mastering Physics understands that grinding away practicing exams doesn't improve your understanding of the material; it just improves your ability to pass the arbitrary evaluation criteria. It's the same with practicing leetcode before interviews. Doing yet another dynamic programming practice problem doesn't really make you a better SWE.

Minmaxing grades and other external rewards is how we got to the place we're at now. Please stop enshittifying education further.

Also, with all the progress in video gen, what does recording the webcam really do?

Ask about any teacher, scalability is a serious issue. Students being in classes above and under their level is a serious issue. non-interactive learning, leading to rote memorization, as a result of having to choose scaling methods of learning is a serious issue. All these can be adjusted to a personal level through AI, it's trivial to do so, even.

I'm definitely not sold on the idea of oral exams through AI though. I don't even see the point, exams themselves are specifically an analysis of knowledge at one point in time. Far from ideal, we just never got anything better, how else can you measure a student's worth?

Well, now you could just run all of that student's activity in class through that AI. In the real world you don't know if someone is competent because you run an exam, you know if he is competent because he consistently shows competency. Exams are a proxy for that, you can't have a teacher looking at a student 24/7 to see they know their stuff, except now you can gather the data and parse it, what do I care if a student performs 10 exercises poorly in a specific day at a specific time if they have shown they can do perfectly well, as can be ascertained by their performance the past week?

This sounds as though it was written by an LLM too.

On the other hand, can an AI exam really simulate the conditions necessary for improving at this skill? I think this is unlikely. The students' responses indicate not a general lack of expertise in oral communication but also a discomfort with this particular environment. While the author is making steps to improve the environment, I think it is fundamentally too different from actual human-to-human discussion to test a student's ability in oral communication. Even if a student could learn to succeed in this environment, it won't produce much improvement in their real world ability.

But maybe that's not the goal, and it's simply to test understanding. Well, as other commenters have stated, this seems trivially cheatable. So it neither succeeds at improving one's ability in oral communication nor at testing understanding. Other solutions have to be thought of.

This is to be expected. The big commercial LLMs generally respond with text that agrees with the user.

> But here's what's interesting: the disagreement wasn't random. Problem Framing and Metrics had 100% agreement within 1 point. Experimentation? Only 57%.

> Why? When students give clear, specific answers, graders agree. When students give vague hand-wavy answers, graders (human or AI) disagree on how much partial credit to give. The low agreement on experimentation reflects genuine ambiguity in student responses, not grader noise.

The disagreement between the LLMs is interesting. I would hesitate to conclude that "low agreement on experimentation reflects genuine ambiguity in student responses." It could be that it reflects genuine ambiguity on the part of the graders/LLMs as to how a response should be graded.

I could accept this for a 300 students class, but 36? When I got my degree, ALL exams had an oral component, usually more than 30 minutes long. The prof and one or two TAs would take a couple days and just do it. For 36 students it’s more than doable. If I was a student being examined by an LLM I would feel like the professor didn’t care enough to do the work.

LLM oral exams can provide assessment in a student's native language. This can be very important in some scenarios!

Unlimited attempts won't work in the presented model. No matter how many cases you have, all will eventually find their way to the various cheating sites.

There is no silver bullet. There's no solution that works for all schools. Strategies that work well for M.I.T. with competitive enrollment and large budgets won't work for a small community college in an agricultural state, with large teaching loads per professor, no TAs, and about 15-25 hours of committee or other non-teaching work. That was my situation.

Teaching five courses and eight sections, 20-30 students per section, 10-20 office hours every week (and often more if the professor cared about the students), leaves little time for grading. In desperation I turned to weekly homework assignments, 4-6 programming projects, and multiple choice exams (containing code and questions about it). Not ideal by any means, just the best I could do.

So I smile now (I'm retired) when I hear about professors with several TAs each, explaining how they do assessment of 36 students at a school with competitive enrollment.

Part 2 is that when you are ready, an examiner sits with you, looks over your stuff and asks questions about it, like clarifications, errors to see if you can fix them, fake errors to see if you can defend your solution, sometimes even variations or unrelated questions if they are on the fence as to the grade. Typically that takes 3-10 minutes per person.

Works great to catch cheating between students, textbook copying and such.

Given that people finish asynchronously you don't need that many examiners.

As to being more stressful for students I never understood this argument. So is real life.. being free from challenge based stress is for kindergarteners

In a presentation, you are in control. You decide how you will present the information and what is relevant to the theme. Even if you get questions, they will be related to the matter at hand that you need to dominate in order to present.

In oral exams, the pressure is just too great. I doubt it translates to a proper job. When I'm doing my job, I don't need to come up with answers right there on the spot. If I don't remember something, I have time to think it through, or to go and check it out. I think most jobs are like this.

I don't mind the pressure when something goes wrong in the job and needs a quick fix. But being right there, in an oral exam, in front of an antagonistic judge (even if they have good intentions) is not really the way to show knowledge, I think.

(I invented some kind of metric based on a centered gaussian around a country ahaha)

One big issue that I had is that the system asked for a number in dollars, but if I answer $2000,2000,2000 per agent per month, the answer was always the same, I cannot accept a number, give it in words, after many tries I stopped playing, it wasn't clear what it wanted.

I could see myself using the system. With another voice as it was kind of agressive. More guidelines would be needed to know exactly how to pass a question or specify numbers.

I don't know my grade, so I don't know how much we can bullshit the system and pass

The student is supposed to submit a whole conversation with an LLMs.

The LLM is prompted to answer a question or resolve a problem, and the LLM is there to assist. The LLM is instructed to never reveal the answer.

More interesting is the concept that the whole conversation is available to the instructor for grading. So if the LLMs makes mistake, or give away the solution, or if the student prompt engineer around it. It is all there and the instructor can take the necessary corrective measures.

87% of the students quite liked it, and we are looking forward to doubling the students that will be using it next quarter.

Overall, we are looking for more instructor to use it. So if you are interested in it please get in touch.

More info on: https://llteacher.blogspot.com/

The key implementation detail to me is that the whole class is sitting in on your exam (not super scalable, sure) so you are literally proving to your friends you aren’t full of shit when doing an exam.

I wonder: with a structure like this, it seems feasible to make the LLM exam itself available ahead of time, in its full authentic form.

They say the topic randomization is happening in code, and that this whole thing costs 42¢ per student. Would there be drawbacks to offering more-or-less unlimited practice runs until the student decides they’re ready for the round that counts?

I guess the extra opportunities might allow an enterprising student to find a way to game the exam, but vulnerabilities are something you’d want to fix anyway…

https://sibylline.dev/articles/2025-12-31-how-agent-evals-ca...

I would prefer to write responses to textual questions rather than respond verbally to spoken questions in most cases.

Wouldn't a written exam--or even a digital one, taken in class on school-provided machines--be almost as good?

As long as it's not a hundred person class or something, you can also have an oral component taken in small groups.

> 36 students examined over 9 days > 25 minutes average (range: 9–64)

It appears that they examined only 4hrs each day, one student at a time. This is incredibly inefficient.

In my experience, the greatest benefit of doing something like this would be to be able to run these exams in parallel, while retaining a somewhat impartial grading system.

> The grading was stricter than my own default. That's not a bug. Students will be evaluated outside the university, and the world is not known for grade inflation.

Good!

> 83% of students found the oral exam framework more stressful than a written exam.

That's alright -- that's how life goes. This reminds me of a history teacher I had in middle school who told us how oral exams were done at the university he had studied in: in class, each student would come up to the front, pick three topics at random from a lottery-ball-picker type setup, and then they'd have a few minutes in which to explain how all three are related. I would think that would be stressful except to those who enjoy the topic (in this case: history) and mastered the material.

> Accessibility defaults. Offer practice runs, allow extra time, and provide alternatives when voice interaction creates unnecessary barriers.

Yes, obviously this won't work for deaf students. But why must it be an oral examination anyways? In the real world (see above example) you can't cheat at an oral examination because you're physically present, with no cheat sheets, just you, and you have to answer in real time. But these are "take-at-home" oral exams, so they had to add a requirement of audio/video recording to restore the value of the "physically present" part of old-school oral exams -- if you could do something like that for written exams, surely you would?

Clearly a take-home written exam would be prone to cheating even with a real-time AI examiner, but the real-time requirement might be good enough in many cases, and probably always for in-class exams.

Oh, that brings me to: TFA does not explicitly say it, but it strongly implies that these oral exams were take-at-home exams! This is a very important detail. Obviously the students couldn't do concurrent oral exams in class, not unless they were all wearing high quality headsets (and even then). The exams could have been in school facilities with one student present at a time, but that would have taken a lot of time and would not have required that the student provide webcam+audio recordings -- the school would have performed those recordings themselves.

My bottom-line take: you can have a per-student AI examiner, and this is more important than the exam being oral, as long as you can prevent cheating where the exam is not oral.

PS: A sample of FakeFoster would have been nice. I found videos online of Foster Provost speaking, but it's hard to tell from those how intimidating FakeFoster might have been.

https://ednutting.com/2025/11/25/return-of-the-viva.html

---

> Only 13% preferred the AI oral format. 57% wanted traditional written exams. 83% found it more stressful.

> Here is an email from a student: "Just got done with my oral exam. [...] I honestly didn't feel comfortable with it at all. The voice you picked was so condescending that it actually dropped my confidence. [...] I don't know why but the agent was shouting at me."

> Student: "Can you repeat the question?" Agent: paraphrases the question in a subtly different way.

> Students would pause to think, and the agent would jump in with follow-up probes or worse: interpret the silence as confusion and move on.

---

Based on these highlights, you'd think the experiment was a wash. The author disagrees!

> But here's the thing: 70% agreed it tested their actual understanding: the highest-rated item.

Man, you could shoot me with a gun, then make me write an essay, & I'd be forced to agree that you had tested my "actual understanding." That doesn't mean my performance wouldn't suffer. Also, 70% is not very high. That's barely two thirds.

Even the grading was done by LLMs (rather than having a TA grade a transcript, and the results were lower. The author defends this by saying, "Students will be evaluated outside the university, and the world is not known for grade inflation," but the world isn't "known for grade inflation" because it doesn't grade you at all. That's not even an excuse, it's just nonsense. It'll toughen you up, or whatever. Was this post written by an LLM too?

> Take-home exams are dead. Reverting to pen-and-paper exams in the classroom feels like a regression.

"Regression"? I mostly wrote pen & paper exams, and I only graduated a few years ago. If students want more flexibility, team up with other courses to supervise multiple exam sessions. Leaked questions aren't going to be any more of a problem than it was for take-home exams, especially since they can't take the booklets with them when they go.

It sounds like these students had a terrible time, and for what? Written exams work fine. These guys just wanted to play with LLMs.

...but OTOH if cheating is so easy it's impossible to resist and when everyone cheats honest students are the ones getting all the bad grades, what else can you do?

Thank god I had a chance to study in pre-AI times.

First, the business school administration and faculty firmly commits, that plagiarism, including with AI, means prompt dismissal.

Then, the first time you have a suspicion of plagiarism, you investigate.

After the first student of a class year is found guilty, and smacked to curb, all the other students will know, and I bet your problem is mostly solved for that class year.

Then, one coked-up nepo baby sociopath will think they are too smart or meritorious to "fail" by getting caught. Bam! Smacked to the curb.

Then one of those two will try sue, and the university PR professionals will laugh at them, for putting their name in the news as someone who got kicked out of business school for cheating. The business school will take this opportunity to bolster their reputation for excellence.

At this point, it will become standard advice for the subsequent class years, that cheating at this school is something only an idiot loser does, not a winner MBA.

Some people dream that technology (preferably duly packaged by for-profit SV concerns) can and will eventually solve each and every problem in the world; unfortunately what education boils down to is good, old-fashioned teaching. By teachers. Nothing whatsoever replaces a good, talented, and attentive teacher, all the technologies in the world, from planetariums to manim, can only augment a good teacher.

Grading students with LLMs is already tone-deaf, but presenting this trainwreck of a result and framing it as any sort of success... Let's just say it reeks of 2025.