My intuition is that there must be some relatively simple ways to tweak the format of assignments or classes so that it's harder to use tools like ChatGPT to cheat.
Yep, that is how my degree was graded, a UK university. It was a distance learning course, but you had to physically attend an exam center and write your answers with pen and paper, no access to phones or computers.
The basic problem is that most classwork is plowing the same ground over and over. "Generative AI" is really a glorified autocomplete--anything that's done over and over and over it will become able to do. You want to stop AI, you have to give the students novel assignments.
I started suggesting this a few years ago, but to exploit their agreeableness, pre-write or have an LLM generate a paper for the students to correct. Additional points are awarded for expanding details and content or contributing improvements to formatting and style.
This allows students to express their individual strengths while ensuring they also know how to properly proofread.
Also, dock points for incorrect / non-existent citations. Alternatively, pre-select and provide sources without a topic, and let the students assemble a paper from them with no length requirement, but a citation minimum. While this may eventually be overcome with RAG and summarization, it's a weakness of current models.
Higher education is destroyed by standardized testing. If you force every square peg to fill a round hole, there are going to be some people that go to absurd lengths to get a passing or perfect grade. It's been like this for a while now.
AI cheating is really the tip of the iceberg, here. High school and college educations heavily rely on cookie-cutter assessment programs that aren't just unreliable but also easily cheated on. The cynic in me says that it's the College Board's fault for doubling-down on bubbling tests and SAT scores when today's graduates all cheat their way through them anyways.
My Students are allowed to do theyr homework with ai. But it's not theoretical assignments, but doing projects. I assume that being able to use AI in the future work is a important part, so they should use it early and experiment with it during study. Just because calculators got invented math didnt just vanish, they had to adapt what to teach and how to do assignments.
The ones who struggle the most are the ones that do the same assignments since 20 years and now are confronted with a new tool that makes theyr tests obsolete.
Yup. I've been programming for 40 years. The toolset has gotten vastly better than it was in the old days but it's always been about letting the system take care of the scutwork. The oft-repeated claim that such-and-such a tool will mean people can do their own stuff never pans out--no, computers excel at scutwork but you still need the human to lay out the picture.
However, you say "tests"--no, the problem is assignments, not tests.
I teach math in the first year of the university. Stuff like solving a 2×2 or 3×3 system of linear equations or the derivative or integral of x*exp(x^2).
They are very important for the phisics or whatever class next year, but they are just tools too boring to make a special project about them.
I had a math instructor in my small, unknown technical school for my mechanical engineering program. He took a huge amount of initiative to work with the profs in the engineering department to make our assignments into word problems showing us exactly how what we were learning would be used later on.
I think it was a lot more useful because we could see how each of the subparts in a multi-part question worked together to solve engineering problems.
I know this isn't possibly in a generalist class with students from many departments, but there are some ways to make them less boring.
I'd recommend researching where such math is applied in the industry and at least show them where it's used. I have rarely understood why I should learn something. While having to research it again years later.
Matrix multiplications are for example used in game engines or in image processing. One example of rather complex algorithm is potential based interception algorithms for moving objects (Ai for Game Developers - O'Reilly).
The problem is that many tests aren't about asking the specific questions; those questions are mere proxies for the presence of foundational knowledge. They're testing whether a student is building up a rich mental model of the domain.
So we're not really testing students to see if they can do X times Y arithmetic problems, and we're not anticipating that future work involves being a calculator. Instead we're using these tests as proxies for whether a student has a good model of multiplication.
ML tech has the possibility of making someone look like they have the foundational knowledge. We can ask people to design "better" tests, but that also makes things harder to cope for everyone else! Must we all take increasingly difficult and fatter tests when a cheap proxy would've sufficed?
Also, perhaps some might say that foundational knowledge is not foundational because ML tech can one day do everything. Yes, this is a possibility. That one day there's no point in learning foundational math because ML can handily beat people just like it does in Go and Chess. But in such a world we won't be talking about students cheating in school, we'll be talking about massive economic revolution.
On the one hand, "good". Higher education needs a rethinking. ROI is all out of whack for too many students. On the other hand, calculators didn't destroy math education. We're probably not more than 3 years away from some form of AI being better than 80 percent of teachers for most grade levels and subjects.
I sure hope you are right (but i dont thonk you are) k12 costs are the second biggest line item on most provimces bidgets after healthcare and way more than half of that is salaries. being able to fire 80 to 90 % of those jobs would be amazing.
All assignments should be done in class with paper and pencil or locked-down Chromebooks. No phones. Lectures can be put on youtube for viewing outside of class. I had a couple math classes like this and it worked well.
A lot of people seem to misunderstand the purpose of tests in school.
When a test for a calculus class asks you to find a particular integral or a test for a music theory class gives you a melody and asks you to add two more voices to produce a three voice fugue following the voice leading and counterpoint rules that would be used by typical Baroque composers the professor is not asking because the calculus professor actually wants to know value of the integral or the music theory professor actually needs to have a fugue written around that melody.
They are asking because they were supposed to have taught you how to evaluate such integrals or apply the rules of Baroque fugue composition, respectively, and they want you to demonstrate that you have learned that.
Doing this requires that you solve the problem, not that you get some tool to do it.
Yes, after school you will use such tools to do most such problems that come up but that's completely irrelevant because at work you are not being asked to solve those problems to demonstrate you know how to solve those problems manually. You are being asked to solve them because you actually need the solution.
> Doing this requires that you solve the problem, not that you get some tool to do it.
Ideally, yes. But this reveals the problem at the heart of busywork assignments and standardized testing; there is more than one way to skin a cat. You're still graded by an objective set of qualities that can be gamed and manipulated to make yourself stand out against more competent students. You probably shouldn't just copy a calculator or the circle of fifths for your homework assignment; but if it fulfills the demands of the question, you might as well.
You have to award points for what you want to see. If you grade for completion, you will get minimum-effort work back from your students. If you award points for writing out computation by-hand, you'll get more procedure-oriented students. Testing fails students because it inherently forces them to look at their grade from a results-oriented perspective.
An easy way to address it just project based grading. But of course, this needs a lot more graders than just a single professor to work with large class sizes. Which means more money. And it's hard to get money for education in the US.
So, it's not really AI that's at fault, as with any new technology, it disrupts; That's what new things do. It's just that the status quo is this lazy approach to education that hamstrings teachers by denying them the resources they need to do the work well, in the name of saving money.
An ironic twist on this "saving money" is that spending more on people, nets more returns. Educate someone well, and they can do high quality work, which makes a lot more money than minimal investment. Spend money on healthcare and prevention, and medical costs overall go down, as large problems are addressed when they're still small.
But I guess the lure of "lower taxes" is enough for this sort of thing to persist.
32 comments
[ 4.9 ms ] story [ 69.5 ms ] thread(Of course the people that brought us the Scantron think that's too much work, but the reality is a normal institute can grade 500 exams in a weekend)
This allows students to express their individual strengths while ensuring they also know how to properly proofread.
Also, dock points for incorrect / non-existent citations. Alternatively, pre-select and provide sources without a topic, and let the students assemble a paper from them with no length requirement, but a citation minimum. While this may eventually be overcome with RAG and summarization, it's a weakness of current models.
AI cheating is really the tip of the iceberg, here. High school and college educations heavily rely on cookie-cutter assessment programs that aren't just unreliable but also easily cheated on. The cynic in me says that it's the College Board's fault for doubling-down on bubbling tests and SAT scores when today's graduates all cheat their way through them anyways.
But the worst part is that you cannot even say whether one pedagogical approach is better than another because measurements are now incommensurate.
If that is the case, then the employers dependent on that workforce need to support it.
Perhaps AI is teaching industry that they can no longer depend on a free lunch.
My Students are allowed to do theyr homework with ai. But it's not theoretical assignments, but doing projects. I assume that being able to use AI in the future work is a important part, so they should use it early and experiment with it during study. Just because calculators got invented math didnt just vanish, they had to adapt what to teach and how to do assignments.
The ones who struggle the most are the ones that do the same assignments since 20 years and now are confronted with a new tool that makes theyr tests obsolete.
However, you say "tests"--no, the problem is assignments, not tests.
They are very important for the phisics or whatever class next year, but they are just tools too boring to make a special project about them.
I think it was a lot more useful because we could see how each of the subparts in a multi-part question worked together to solve engineering problems.
I know this isn't possibly in a generalist class with students from many departments, but there are some ways to make them less boring.
Matrix multiplications are for example used in game engines or in image processing. One example of rather complex algorithm is potential based interception algorithms for moving objects (Ai for Game Developers - O'Reilly).
So we're not really testing students to see if they can do X times Y arithmetic problems, and we're not anticipating that future work involves being a calculator. Instead we're using these tests as proxies for whether a student has a good model of multiplication.
ML tech has the possibility of making someone look like they have the foundational knowledge. We can ask people to design "better" tests, but that also makes things harder to cope for everyone else! Must we all take increasingly difficult and fatter tests when a cheap proxy would've sufficed?
Also, perhaps some might say that foundational knowledge is not foundational because ML tech can one day do everything. Yes, this is a possibility. That one day there's no point in learning foundational math because ML can handily beat people just like it does in Go and Chess. But in such a world we won't be talking about students cheating in school, we'll be talking about massive economic revolution.
Any teacher who is letting students cheat with ChatGPT is lazy.
Source: I teach for a living.
Source: I teach for a living.
I teach my students to use AI ethically.
I teach them that ethical use of AI is when the AI helps you learn so that when the AI is not there you are more capable.
But if using the AI as a crutch so you do not have to learn things is unethical.
Why should we believe any of it?
> There are a number of platforms that, when used effectively, can give instructors a clear view of how students are using available AI tools.
I can't help but think his company's product is one of those.
And that it requires installing spyware on the computers students use for their assignments.
Schools should account for LLM use. In fact, schools should have an official LLM or sanctioned 3rd party LLMs that they allow students to use.
When a test for a calculus class asks you to find a particular integral or a test for a music theory class gives you a melody and asks you to add two more voices to produce a three voice fugue following the voice leading and counterpoint rules that would be used by typical Baroque composers the professor is not asking because the calculus professor actually wants to know value of the integral or the music theory professor actually needs to have a fugue written around that melody.
They are asking because they were supposed to have taught you how to evaluate such integrals or apply the rules of Baroque fugue composition, respectively, and they want you to demonstrate that you have learned that.
Doing this requires that you solve the problem, not that you get some tool to do it.
Yes, after school you will use such tools to do most such problems that come up but that's completely irrelevant because at work you are not being asked to solve those problems to demonstrate you know how to solve those problems manually. You are being asked to solve them because you actually need the solution.
Ideally, yes. But this reveals the problem at the heart of busywork assignments and standardized testing; there is more than one way to skin a cat. You're still graded by an objective set of qualities that can be gamed and manipulated to make yourself stand out against more competent students. You probably shouldn't just copy a calculator or the circle of fifths for your homework assignment; but if it fulfills the demands of the question, you might as well.
You have to award points for what you want to see. If you grade for completion, you will get minimum-effort work back from your students. If you award points for writing out computation by-hand, you'll get more procedure-oriented students. Testing fails students because it inherently forces them to look at their grade from a results-oriented perspective.
So, it's not really AI that's at fault, as with any new technology, it disrupts; That's what new things do. It's just that the status quo is this lazy approach to education that hamstrings teachers by denying them the resources they need to do the work well, in the name of saving money.
An ironic twist on this "saving money" is that spending more on people, nets more returns. Educate someone well, and they can do high quality work, which makes a lot more money than minimal investment. Spend money on healthcare and prevention, and medical costs overall go down, as large problems are addressed when they're still small.
But I guess the lure of "lower taxes" is enough for this sort of thing to persist.