Post

Conversation

This is an extraordinary chart. If you give take-home exams in 2026, you are sabotaging honest students.
Quote
Brendan Nyhan (@BrendanNyhan on 🟦☁️)
@BrendanNyhan
Friends don't let friends continue to give exams and papers that are not adapted to the current AI landscape (this is AFTER 27 bailed so the reality is even worse) PS Props to students 1, 22, and 31
Image
David Watson 🥑
Post your reply

Amazing. I just went through 3 years of law school and never wrote an exam or a paper with AI. I'm a sucker. Seems like the future is either in person exams or essays with frequent periods to hand in work in progress. Or computers that lock out AI.
I definitely used AI in my advanced legal research take-home that allowed AI, but I also took the opportunity to write within the final that the AI could now do ~all of it without assistance and it was worth reconsidering approach moving forward
Do you think if schools were serious about catching and appropriately punishing cheaters this would still be an issue? Removing take home tests that would otherwise have been pedagogically appropriate is also sabotaging them.
Yes. It is structurally impossible, I think, to design a take-home exam on which you can reliably restrict AI and catch cheaters, and there are extremely strong incentives to cheat. Maybe Pangram can prove me wrong, but I think the format has been ruined.
Even before LLMs became competent, take-home exams usually rewarded the students willing to work the entire take-home period over those who put a reasonable amount of time into the test and moved on.
yeah, I hate take-home for that reason. “I can make it harder because you have unlimited time”—no! bad! give me a three-hour block and an amount of work that will take three hours of focus, please
punishes honest students and rewards cheaters, or if LLMs are allowed, punishes students who want to learn and rewards students who want to prompt
Problem sets retain their correct (imo) role, which is “train the students in preparation for a graded final.” If someone wants to skip out on training, ok, let’s see how they do on a test.
instructors will need to tell students what they need to know and how in depth, then let them use course resources or any other resources to learn it, then come into class and demonstrate they know it, either written or spoken etc. The goal is information understanding, how they
Not only that, you aren't giving useful feedback to that chunk of students who give in to the temptation to cheat, and quite possibly wouldn't cheat absent the temptation. Just make it easier and do it in class. Even before AI, I never liked take home exams. Everyone gets approx.
As a teacher, I spend a lot of time catching AI cheating and thinking of ways to root it out. I'm pretty good at it, it is exhausting and limiting. Worse still, lots of admin like those high scores. Many teachers have to work past admin who would prefer the cheating continue.
If your take home exams can be cheated like this, you’re not actually testing knowledge, you’re testing memory
The future model is the past: interrogation and didactic exams where the student is alone before the teacher (and class) and must fight to defend his positions with logic and rhetoric. The remaining herd-humans will have to rely on ChatGPT to know if argument was true.
There needs to be a special reward for the two students who were honestly bad (or average on the curve?) on both exams.
I think this is a slight overstatement. It's probably true for large intro courses serving students majoring in something else, but take-homes can still work in small upper-level majors courses. The key is that you need the students to buy in on actually learning the subject.
Perfect scores on homeworks are strong negative signals at this point, this is clear from being a student
Yeah I'm sure not all of them cheated, some just found the second half harder, but that's stunning
lol I’m sure the last 3 just didn’t take the final or something but it’s very funny to think of a college student sitting down and getting absolutely nothing right.
Square profile picture
90% of your simulated profits can be yours. Take a single step and start earning rewards based on your performance!
The media could not be played.
Yeah. I figured that out in 2022. So I allowed AI with explanation and attribution. I got all AI slop. Now we’re in-class, paper-based.
I feel the same way about take home interview coding questions, where they say "don't use AI"... I just told them no thanks. How is that a filter?